Bringing Capslock analysis to

Jess McClintock and John Dethridge, Google Open Source Security Team

Security usability is hard — security best practices often add further toil on developers, to the point where usability and security are often considered as direct tradeoffs. But they don’t have to be. There are many fantastic analysis tools and platforms for developers who want to audit their transitive dependencies or vendor a third party package. We aim to make this process less involved by running analyses centrally and making results directly available to open source consumers without any additional steps for maintainers.

We are excited to have added Capslock results for Go packages. Capslock is a Google open source tool that identifies “capabilities” that packages have — for example, the ability to read files, or to send and receive data on the network. This will provide added visibility into the behaviors of Go packages, to assist in choosing appropriately scoped dependencies, understanding what packages are doing under the hood and noticing when dependency updates require more powerful capabilities. Our aim here, as with as a whole, is to help those choosing and using open source packages to make well informed selections based on as much security-relevant data as possible.

A difference that makes a difference?

Our analysis found that less than 2% of version updates for packages will introduce a new capability requirement. This makes sense, since the set of capabilities required by a package will usually be established by its initial behavior, and if a dependency adds new capabilities, that can be a signal for maintainers that the change is more interesting (or, perhaps, the change warrants further investigation).

We divide these results according to whether the standard library was called directly from the analyzed package or via a transitive dependency. Some capabilities, such as those involving the network, are more likely to be used directly than via a transitive dependency. Interestingly, a whopping 9% of Go packages have a transitive dependency using os/exec!

Capability Analysis for Go

Capslock was launched last year as a CLI to analyze the callpaths of Go packages and report on the privileged capabilities that are accessible. If a package unexpectedly adds a new capability in an update, or uses capabilities that aren’t required for its operation, this can be a risk indicator about the codebase. Capabilities can also be a useful signal for verifying that a package does what is expected — a well designed package should itself require minimal privileges, and allow objects and interfaces like those in io/fs to be passed in that encapsulate privileged capabilities.

Capslock results on include the list of capabilities that have been identified in each particular version of a package. Note that the analysis results are build specific, so functions that are only included when a package is built for a specific operating system might not appear in our reports. You can find more details about this analysis in the Capslock documentation.

Today we are launching Capslock capability results for Go packages on This will be our first step in making capability data more widely available to open source consumers.

Introducing artifact URLs in the Query endpoint

Josie Anugerah, the Open Source Insights Team

Today we’re adding artifact URLs to our v3alpha/Query endpoint. Since the API was launched in April 2023, has supported mapping artifacts to package versions using content hashes.

From user feedback, we discovered that some of the results were surprising. To address this, we’re providing more information about why a hash matches a package version through a new artifacts field in the Query endpoint response. This artifacts field contains the URL of the artifact the hash was calculated from.

This is a breaking change to the v3alpha/Query endpoint. As mentioned in the v3 API blog post, v3alpha is intended for experimental features and v3 for stability. Features that become stable in v3alpha will eventually be added to the v3 API as a non-breaking change. If you have a critical application depending on the API and are still using v3alpha, consider migrating to v3.

New features in the API

Jesper Särnesjö and the Open Source Insights Team

In April 2023, we launched the first public version of the API, v3alpha, complementing the website and BigQuery dataset as a new way of exploring our software supply chain data. Since then, the API has served billions of requests, enabling applications like providing a richer view of SBOM data in GUAC or reporting dependency licenses in OSV-Scanner. Today, we’re adding a few of our most frequently requested features, along with a new version of the API that comes with a stability guarantee.

A new stable version

When we launched the first public version of the API, we named it v3alpha to indicate that it might change over time. In practice, we haven’t needed to make any incompatible changes to it since launch. Today, we’re formalizing our approach to API stability.

First, we’re launching a new version, v3, that comes with a stability guarantee: we will never make incompatible changes to it. It also comes with a deprecation policy: if and when we deprecate the API in favor of a newer version, we will give at least 6 months notice. Deprecations will be announced on this blog, on our documentation site, and on our GitHub repo. The v3 API has almost exactly the same structure as v3alpha has had, with only a few small tweaks to fix inconsistencies and set it up for planned—compatible—changes, and is also available via gRPC or HTTP. We recommend that most users migrate to it.

Second, with v3 providing a stable option, our existing v3alpha API can be a bit more experimental. This means that most new features will start out in the v3alpha API, possibly change in incompatible ways in response to user feedback, and eventually graduate to the v3 API.

Batch request support

First among the experimental features we’re adding to the v3alpha API is a highly requested one: batch support. Many applications currently require making hundreds of API requests for a single user task—such as fetching licenses and security advisories for all your dependencies—which can be cumbersome. With our new GetVersionBatch and GetProjectBatch endpoints, you can instead make a single request containing a batch of identifiers, and get the results in a paginated response.

Purl support

The next new experimental feature is also one requested by users: support for fetching package- or version-level data by purl. Purls, or package URLs, are widely used identifiers for packages and versions from a number of ecosystems, including the six that we currently support. By adding purl support to the API, we hope to lower the barrier to integrate it into existing workflows. Combining our two new features, we’re also adding an endpoint for requesting batches of versions by purl.

But wait, there’s more!

In addition to these brand new features, we’re also filling a few gaps. You can now find packages with similar names, previously launched on the website, using the GetSimilarlyNamedPackages endpoint in the v3alpha API. We’ve made OSS-Fuzz data available from the GetProject endpoint in both the v3 and v3alpha APIs. We’ve also added more data showing how package versions and projects relate to each other, whether by metadata or by SLSA attestations, so users can decide whether the link is trustworthy for their application. SLSA attestations are now also verified by us.

To get started using any of the new API features mentioned in this post, check out the documentation, or visit our GitHub repo for code examples and the gRPC service definition.

One set of requirements, zillions of SBOMs

Francois Galilee and Laurent Simon, Open Source Insights Team

In this post, we explain how dependency resolution works in package managers, with the npm ecosystem as an example. We also explain how it directly affects the accuracy of SBOMs you generate and ingest.

Let’s take the package d3 as an example. Version d3@7.8.5 requires 30 dependencies in its npm package.json. How many different Node.js applications respecting these 30 requirements may be generated? One? Quite a few actually: at the time of writing, there are ~1.9X10^81. That makes as many possible SBOMs for this application as the estimated number of atoms in the universe.

Composition in package managers

Open source ecosystems in general thrive on sharing and reusing components. When a developer builds an application, they compose their application’s code with multiple components created by other developers, that in turn may rely on multiple components created by other developers. To facilitate this composition, each ecosystem provides tools to install libraries and applications.

Among them and of particular interest to us is the dependency resolver, that goes beyond direct dependencies and ensures that transitive requirements are satisfied. For example if an application A depends on two libraries B and C that both in turn depend on a library D with conflicting requirements, which version of D should be chosen?

A diagram containing a package A with a requirement of 1 on packages B and C, which in turn have requirements 0-3 and 2-5 on package D respectively.
A diagram containing a package A with a requirement of 1 on packages B and C, which in turn have requirements 0-3 and 2-5 on package D respectively.

Let’s dive deeper. We will focus our discussion on npm as a concrete example in the rest of the post. The insights are typically applicable to other ecosystems.

Dependency requirements

In the npm ecosystem, npm registry is the de facto public registry and npm the de facto associated tooling. CLI alternatives to npm exist, including yarn and pnpm.

Developers express their dependency requirements in a manifest file “package.json”, specifying a package and a set of acceptable versions for each package. For example, let’s create an application with four dependencies:

$ cat >package.json <<EOF
  "name": "example",
  "version": "0.0.0-alpha.0",
  "dependencies": {
    "d3-time": "^3",
    "d3-array": "<2",
    "array": "npm:d3-array@^2",
    "color": "npm:d3-color@^3"

This manifest declares four dependencies named “d3-time”, “d3-array”, “array”, and “color”, each with a corresponding version range. For example, “d3-time” is declared with the constraint “^3” which means that any 3.x version may be installed. (More information about version declarations are in the official documentation). In addition to version set requirements, packages can be aliased, as illustrated by “array” and “color” that are aliasing “d3-array” and “d3-color” respectively.

Dependency resolution

Given this set of dependency requirements, the ecosystem tooling selects and installs the versions that it deems adequate (often the latest version of the matching set) to create an application. This process is referred to as “dependency resolution”. The selected packages are physically installed as a file tree for npm, so that Node.js imports them at runtime. Below is an example of an installation by npm@6.14.13 for the former manifest:

$ npm -v && npm install && npm ls
+-- array@npm:d3-array@2.12.1  // d3-array@2.12.1 installed under an alias.
| `-- internmap@1.0.1
+-- color@npm:d3-color@3.1.0
+-- d3-array@1.2.4  // d3-array@1.2.4 installed at the root.
`-- d3-time@3.1.0
  `-- d3-array@3.2.4  // d3-array@3.2.4 installed locally, shadowing the root.
    `-- internmap@1.0.1 deduped

In the snippet above, we note that “d3-array” is installed multiple times with different versions:

  • Once under the alias “array” for the main application’s code, as “d3-array@2.12.1”
  • Once for the main application’s code, as “d3-array@1.2.4”
  • Once for “d3-time”, at version “d3-array@3.2.4”. This means “d3-time” will use version 3.2.4.

If we use different tools to install from the same manifest file, they may install different dependency versions. On our example manifest file, npm, yarn, and pnpm produce different installations:

A resolution result for the manifest above using npm 6.14.13 or at version 9.8.1 hoisted.
A resolution result for the manifest above using npm 6.14.13 or at version 9.8.1 hoisted.
A resolution result for the manifest above using yarn 1.22.10 or npm 9.8.1 nested.
A resolution result for the manifest above using yarn 1.22.10 or npm 9.8.1 nested.
A resolution result for the manifest above using pnpm 8.6.9.
A resolution result for the manifest above using pnpm 8.6.9.

All three dependency resolutions are valid and are among the set of graphs that satisfy the constraint requirements. Other factors besides the tool chain can affect the resolution. For example, the time: if we install this manifest today, the result may be different from the installation we made yesterday: version resolutions change as new (dependency) package versions become available or deleted from the npm registry.

Downstream consumers control the composition

It is worth noting that the composition (aka version resolution) is triggered by the downstream user who creates an application. In other words, dependency packages (libraries) are oblivious to the composition. They express requirements for their own direct dependencies as abstract strings (like “d3-time@^3”) that define the set of functionally compatible versions. But it is the tooling run by the downstream user creating the application that selects and installs a concrete version from among this set. The selected version may be different from the version that would have been selected by the dependency package itself, as it is made in a different context, using different tooling, at a different time. For example, d3-array@3.2.4 resolves differently in two different contexts:

The dependencies of d3-array 3.2.4 as a stand-alone library and within an application. In both cases, the dependencies were resolved by npm 9.8.1 using the hoisted strategy, but produce a different set of dependencies.
The dependencies of d3-array 3.2.4 as a stand-alone library and within an application. In both cases, the dependencies were resolved by npm 9.8.1 using the hoisted strategy, but produce a different set of dependencies.

It is impossible for the maintainer of a package to enforce dependencies’ versions in downstream users’ applications. For example, they may try to “pin” their own dependency requirement to a specific version, in the hope to force downstream users to use that particular version transitively. But this still can be overruled by a downstream user by:

  • Using an overrides directive to override the dependency.
  • Bundling of packages.
  • Using a custom alias. In the example manifest, if “d3-array” was defined as an alias for “d3-color”, npm would install d3-color in lieu of d3-array to the surprise of the library.

Furthermore, pinning a dependency (by a strict requirement or by providing a bundle) in a library is considered bad practice because it prevents downstream users from upgrading the dependency independently if they need to (to resolve a vulnerability for example).

So how does all this affect SBOMs? Read on!

Impact of composition on SBOMs

A Software Bill of Materials (SBOM) is a document listing the dependencies (and their relationship to one another) used to build software. For more details, see the Minimum Elements For a Software Bill of Materials (SBOM).

As we have seen in the first part of this blog post, packages that are libraries have dependencies, but their versions are resolved by the final application, not the library itself. When an SBOM is generated by the library maintainers at the time of publication, the dependency resolution happens in a different context from the context in which the final application is built (package manager CLI version, available packages on registry, etc). As a result, the dependencies listed in a library SBOM are irrelevant for downstream applications.


In this post, we saw that one set of requirements yields a vast number of applications: the decision on which concrete dependencies are installed lay in the hands of consumers. So library SBOMs cannot list the exact dependency used, but application SBOMs can. Furthermore, the composition involves dependency resolution that relies on complex algorithms. Given the space for error and the nuances of dependency resolution, it might be beneficial to develop tooling to ensure that the application SBOM describes faithfully what has been installed.

Combining dependencies with commit information

Baqiao Liu, Open Source Insights Team

As part of an internship project, we experimented with finding dependencies that are both important and have few maintainers, based on their public source code repositories and dependency graphs.

Modern software development heavily relies on open source libraries to reduce effort and speed up innovation. However, alongside the many benefits, third-party open source libraries can introduce risk into the software supply chain, and modern ecosystems make it easy to end up pulling in tens, if not hundreds of dependencies. Given limited resources, which dependencies should developers focus on to mitigate the risk of supply-chain attacks? Which dependencies might be exposed to more risk of single points of failure? In other words, we want to not only consider the likes of numpy, but also explore the well-hidden 30LoC single-author packages that everyone depends on.

Dependency graphs provides resolved dependency graphs for packages in several ecosystems. These graphs can be complex and can include thousands of direct and indirect dependencies.

A package’s page (for example react) also includes information on the package’s dependents (the open source packages that include the package as a dependency). This information can help us narrow our focus to important packages - those with many direct or indirect dependents. The number of dependents is a good starting metric, and it is already used as part of one prior importance measure for repos called the OpenSSF Criticality Score. But which of these important packages might be susceptible to the risks of a single maintainer?

Dependency graphs are more powerful with more data

Input data: dependency graphs and authorship data
Input data: dependency graphs and authorship data

Let’s look at a hypothetical example dependency graph, and see how both the package dependency graph and contributor commits to an associated repository can be combined to help developers focus on interest-worthy packages in their supply chains. Suppose that all of a developer’s open source software (OSS) dependencies can be mapped to four repos: A, B, C, and D. This is shown graphically in figure (a).

If we only consider the dependency data, C and D are the most important packages in the ecosystem. They have the largest number of dependents. If an attacker were to introduce a vulnerability in either C or D, more packages within the ecosystem will be compromised than if a vulnerability is introduced in A or B. Similarly, if B were to become unmaintained or no longer updated, it could block the adoption of any vulnerabilities fixed in C or D.

In our example case, we consider not only dependency information, but authorship information as well (shown in figure (b)), derived from source code commit information. In this case, a special pattern emerges: repo D is solely authored by Carol, while A, B, and C are all authored collaboratively by Alice and Bob.

When we consider this commit information, repo D becomes quite interesting because it could represent a higher level of risk. All the work of securing repo D, including coordinating security upgrades, falls to a single developer. In general, it is good to have more eyeballs reviewing changes (“Linus’s law”), or to have additional developers performing upkeep.

In other words, repo D seems important because its authorship is unique and has multiple dependents (both A and B). If we were to rank these packages in order of importance for the supply chain, we could say that D > C > B > A. But is there a way to compute this?

Modeling our intuition

Given our intuition, how do we concretely model a ranking of repos when we might have thousands of repos and tens of thousands of authors? To put it in computer science terms, we can map this to defining a scalable node importance score (“node centrality”) and a way to construct a graph using both dependency and authorship data.

Looking at our original dependency graph (figure (a)), imagine a walker is placed on a random node and always travels in the direction of the arrows. The walker randomly chooses an arrow to follow; if there are no arrows to follow, the walker stops. 50% of the time the walker will end up at C and 50% of the time the walker will end up at D. This “random walk” notion of node importance yields C and D being the most important nodes in the graph. Intuitively, C and D are the “most upstream” nodes and attacking them will have the highest impact on the ecosystem, and edges in the graph represent the delegation of security risk and best practices. The more upstream a node is, the more repos have delegated their risks to the node.

Augmented Dependency Graph
Augmented Dependency Graph

Let’s play with the concept of A, B, and C being related to each other because they share their distribution of authors. The natural way to model relationships in a graph is to add edges. We take a simplistic assumption that when repos share authors, they tend to have similar security practices and quality. To model this similarity we add bidirectional edges among all pairs in A, B, and C. This gives us a new graph that not only takes into account dependency information but also authorship information (figure (c)). Reusing our random walker analogy, it is possible for the walker to reach D from any other node, but once our walker is at D, it can no longer travel to another node. We see D as the most important node in the graph: any random walker will eventually land in D with 100% probability and be unable to escape.

The above argument captures our intuition to use the well-known PageRank algorithm as the measure of node importance. PageRank models a walker starting randomly choosing edges to follow in the graph. The more often a node is visited, the higher its importance. By adding shared authorship edges to the dependency graph, PageRank tends to highlight single point of failures in the graph.

Applying to real data

We expanded upon this idea and added weights to the edges in a Python implementation. Then a test run was performed on a sub-ecosystem of our open source usage (a graph of ~500 nodes and ~10k edges). We were able to confirm the general trend: using source code commit data highlights important packages with potential single point of failures better. Let’s take a closer look at a small case study from our analysis that includes the following four packages:

  • golang/protobuf adds support for protocol buffers in Go. Although it has been deprecated, it is still alive and healthy. It has a variety of contributors.
  • josharian/intern is a Go library to store the same strings in the same memory location. It has widespread usage and is largely written by a single person.
  • numpy/numpy is a very popular numerical library for Python. It is active and healthy, with many contributors.
  • google/go-cmp is a utility library to compare values for testing in Go. It is popular and largely written by a single person.

Here are the relative PageRank ranks of these four packages before and after introducing authorship data:

Rank only w/ Dependency Rank w/ Dependency + Coauthorship
golang/protobuf 1 2
josharian/intern 2 1
numpy/numpy 3 4
google/go-cmp 4 3

Table: Ranks of packages before and after introducing shared-authorship data; lower rank means higher relative importance.

Notably, josharian/intern and google/go-cmp have fewer contributors than the other two packages, and thus rank higher when we consider both dependency and co-authorship.

More can be done with provides data to enable developers to perform data-driven decisions to secure their supply chains. On top of, developers and researchers can supply additional data to customize the packages they focus on. We showed that using source code commit data we can additionally identify potential single point of failures in the supply-chain. All of the dependency data mentioned is publicly available via the API and BigQuery datasets, while authorship data can be obtained from the source code repositories associated with the packages.

Open source provides a wealth of data, and we welcome research ideas related to network analysis or general data science that can help unlock new insights about this important resource. If you have any research ideas or feedback, please open an issue or contact us at

This work was performed as part of a Google internship program. If you’re interested in working on open source security, we encourage you to apply to Google’s internship program!

Finding similar package names

Josie Anugerah, Open Source Insights Team

Have you ever seen two packages with similar names and been unsure of which one to install?

Choosing the right package is a security decision. Malicious software supply chain attacks have increased 633% year over year according to Sonatype’s State of the Software Supply Chain report.

Typosquatting involves an attacker uploading a malicious package with a similar name to a popular package in the hopes that users will be confused and download their malicious package instead. Recent high-profile attacks mentioned in the Sonatype report include rustdecimal typosquatting on rust_decimal, and pymafka typosquatting on pykafka and many packages typosquatting on colors.

The team has been thinking of ways to help you find the right package, so we’re excited to launch the similar package names feature on the website! These similarity results are available for three open source ecosystems: npm, PyPI and Cargo.

Screenshot of showing packages with names similar to jost.
Screenshot of showing packages with names similar to jost.

We are currently calculating the following forms of package name similarity:

Packages may have coincidently similar names due to the large number of packages within the ecosystem, but there are other reasons two packages may have similar names:

  • The packages may be intended to be used for different variations of a language (jest is similar to @types/jest)
  • The packages may provide functionality of the same category (rson is similar to bson and json)
  • One package may be named in homage to another more popular package (redrx and redux, jost and jest)
  • A package may be intentionally named to confuse users, an attack category known as typosquatting (reacy is similar to react)

We have found that combining similar name calculations with a popularity metric (such as dependents or downloads) helps to narrow down the noise. In particular, prioritizing similar name pairs by dependents helped to reduce uninteresting results. When we calculated similarity for all possible package pairs, we observed a lot of noise; a lot of packages were similar to each other for no reason other than that there are so many packages in each ecosystem, and so few short or memorable names. For this reason, we calculate similarity only to the most popular packages within each ecosystem, which led to more meaningful results. And to make the results even more useful, we’ve ordered the similar names list by dependents, listing the total dependent count across all versions of each package.

Our new similar names feature is available on npm, PyPI and Cargo. We have computed 12k similar name pairs for npm, 4k for Cargo, and 5.5k for PyPI.

We plan to keep iterating on our similar names calculations and to bring this information into our API and BigQuery datasets in the future. We hope this feature will help you to find the right packages that you are looking for!

Introducing OSS-Fuzz insights

Eve Martin-Jones, Open Source Insights Team is continually adding new features to help developers assess the security of open source projects and the risks posed by adopting them as dependencies. Today, is excited to announce the integration of data from projects tested by the OSS-Fuzz service. This new integration will provide users with a signal that the maintainers of a project are actively maintaining good security practices — including preventive measures — to safeguard the project from major vulnerabilities.

Fuzz testing, or fuzzing, is an automated software testing technique that involves providing random data as input to a program to find bugs that might not be found by other testing methods such as manual or unit testing. To developers considering using an open source project, fuzzing provides a positive signal about the security posture of that project: it shows investment from maintainers and ongoing work to discover and mitigate vulnerabilities.

Google’s OSS-Fuzz is a free service that continuously fuzzes critical open source projects. Its fuzzing runs all day, every day, and an individual project may be fuzzed more than once in a 24-hour period to catch new bugs introduced with code changes as soon as possible. As of July 2023, OSS-Fuzz has helped identify and fix over 9,600 vulnerabilities and 30,600 bugs across more than 1,000 projects, including widely used projects such as netty and spring-framework. now tells you whether a project is fuzzed with OSS-Fuzz and, if so, the percentage of lines of code covered and the configuration details for the project, which show how thoroughly and in what way the project is fuzzed. For example, google/leveldb’s page on shows that 75.1% of the project is fuzzed, and links directly to the configurations in the OSS-Fuzz GitHub repository for those who want to dig deeper into the details of how the project is fuzzed:

Screenshot of OSS-Fuzz UI for the google/leveldb project
Screenshot of OSS-Fuzz UI for the google/leveldb project

We are pleased that this integration will help our users to make more informed security decisions and will highlight the investments that maintainers have made into their project’s security. To get started checking out the fuzzing data for your favorite project, just navigate to the project’s page on!

Enumerating dependents using BigQuery

James Wetter, Open Source Insights Team

Open Source Software (OSS) allows developers to share reusable parts of code across projects, teams and organizations. As a result many thriving ecosystems of interdependent OSS packages have developed. Many OSS packages depend on other OSS packages to function.

We compute a full set of transitive dependencies for each version of each package, and we call this the “dependency graph”. This data is available on our web site, API, and BigQuery dataset. We also compute the inverse of these dependency graphs, providing the full set of versions that depend on any given version, and we call these “dependents”.

The set of packages that depend on a given package is useful for a number of reasons. This blog post demonstrates how to fetch all the dependents of a package within the dataset using BigQuery.

Why do we need dependents?

There are various uses for the set of dependents of a package.

For example, the number of the dependents—direct and indirect—may indicate the level of interest and adoption of a package. Well known popular packages such as react or have tens of thousands of dependent packages published in their respective package management systems. Sorting packages by dependent count can help identify some of the most critical packages within OSS ecosystems.

Additionally, when a vulnerability is discovered in a package its set of dependent packages is highly valuable. It provides insight into the scope of vulnerability across an ecosystem. In some cases the dependents of the affected package may need to act to help propagate a fix through the software supply chain to end users. Access to dependent sets provides a means to identify such packages.

Finally, OSS maintainers can also benefit from being able to identify the many consumers of their packages and better understand how and where their package is used. For example this information may help prioritize future work on their package.

Sample queries

Let’s dive into some BigQuery examples. These samples will select the packages that depend on, but it is easy to adapt them for any other package. Currently the full set of dependents for a given package can only be accessed via BigQuery.

All dependent versions

Our first example fetches a list of all versions of all packages tracked by that depend on the Go package version v3.0.1.

  System = 'GO'
  AND Name = ''
  AND Version = 'v3.0.1';

Currently dependent versions

The previous query fetches all versions of all packages that depend on v3.0.1. Multiple versions of some packages will often be included in the result set. This means counting the number of resulting rows will not correspond to the number of unique dependent packages.

Additionally a package may have required version v3.0.1 at some time in the past, but has since removed or updated its dependency requirement.

To select unique packages that currently depend on version v3.0.1 we can filter the result set to include only the versions that are the newest release of their package.

  System = 'GO'
  AND Name = ''
  AND Version = 'v3.0.1'
  AND DependentIsHighestReleaseWithResolution;

This query can easily be adjusted to find all packages whose highest release depends on any version of

  System = 'GO'
  AND Name = ''
  AND DependentIsHighestReleaseWithResolution;

Direct or indirect dependents only

The result sets returned by all the queries provided so far include both direct and indirect dependents of To find the packages that import directly we can make use of the MinimumDepth column of the Dependents table.

This column contains the minimum depth of the dependency in the corresponding dependency graph. It is a minimum depth because there may be multiple paths to a dependency.

A depth of 1 indicates direct dependency. A depth greater than 1 indicates an indirect dependency.

The following query selects all unique packages that currently depend on any version of directly.

  System = 'GO'
  AND Name = ''
  AND DependentIsHighestReleaseWithResolution
  AND MinimumDepth = 1;

Limitations of these queries

There are some caveats to the queries provided in this post that are worth consideration.

The data does not include closed source dependents

The dataset only includes software that has been published on one of the dependency management systems tracked by Consequently no queries can contain closed source code that depends on a given package.

A package may see wide spread use in proprietary applications, but this popularity will not necessarily be reflected in the number of publicly available dependents.

Context matters

It is common practice across most dependency management systems to allow libraries to specify a range of compatible versions for each of their dependencies. As a result most OSS packages can have their dependency requirements met by many different dependency graphs.

The context in which a library is used can determine the exact dependencies that will be installed.

To compute the dependency and dependent relation in the dataset a single dependency graph is resolved for each version of every package tracked by The dependencies we have should be similar to those obtained when installing the dependencies of a package with native tooling in a clean workspace on a Linux machine. This is also true of API calls that return dependencies like GetDependencies.

It is important to note that different dependency resolutions are possible.

Closing thoughts

Securing the software supply chain is essential, and understanding the complex interrelationships of OSS software is a key part of that.

This post has shown some of the ways the dataset can be used to achieve this goal. We are excited to see what you can do with this dataset.

Introducing NuGet Support for

Josie Anugerah, Open Source Insights Team

Today we’re launching support for NuGet, the .NET package manager. We have 350k NuGet packages and 5.6 million package versions available through our API, website and BigQuery dataset. already supports npm, Go, Maven, PyPI, Cargo and we’re excited to add support for NuGet as another major open source ecosystem. As software supply chain attacks continue to increase in number and in complexity, it’s becoming more important than ever to understand the software that we depend on. We hope that more developers will be able to gain insight into their dependencies through our NuGet support.

Introducing Requirements

NuGet is our first supported ecosystem to feature dependency requirement data instead of dependency graphs.

Requirements are the link between a package and its dependency graph. In NuGet, requirements are specified by including a dependencies tag in the .nuspec or a package reference tag in the .csproj file.

    <!-- Require a version of Castle.Core between 4.0.0 and 5.1.0,
      including 4.0.0 but excluding 5.1.0. -->
    <dependency id="Castle.Core" version="[4.0.0, 5.1.0)" />
    <!-- Require a version of Serilog that's >=2.12.0, preferring lower versions. -->
    <dependency id="Serilog" version="2.12.0" />

Example of a dependencies tag in a .nuspec file. Note that 'dependency' in NuGet means 'requirement'. In we make a distinction between a dependency (a resolved requirement) and a requirement (a package constraint).

These requirements are read by the NuGet resolver and resolved into a dependency graph. For example, the requirements above would resolve into the following graph.

The resolved dependency graph for the .nuspec snippet above. My.Package/1.0.0 depends on Castle.Core/5.0.0 and Serilog/2.12.0, and transitively depends on four more System packages.
The resolved dependency graph for the .nuspec snippet above. My.Package/1.0.0 depends on Castle.Core/5.0.0 and Serilog/2.12.0, and transitively depends on four more System packages.

In the case of NuGet, it’s possible to specify requirements for specific target frameworks.

   <group targetFramework=".NETFramework4.6.2"/>
   <group targetFramework=".NETStandard1.5"">
       <dependency id="Moq" version="[4.7.0, 4.11.0)"/>
       <dependency id="NETStandard.Library" version="1.6.1"/>
   <group targetFramework=".NETStandard2.0">
     <dependency id="Moq" version="[4.7.0, 5.0.0)"/>
     <dependency id="Castle.Core" version="5.1.1"/>

An example .nuspec snippet showing multiple dependency groups for different target frameworks.

Requirements are usually interpreted within the rules of semantic versioning (semver), but the semver standard only covers version numbers. It doesn’t go into how requirements should be specified. So for each ecosystem there are different requirement operators and rules for requirement interpretation. For NuGet, there is official documentation on requirement rules and resolution. You can find a brief comparison of requirement specification rules across ecosystems in our glossary.

Since requirements determine the allowable dependency graphs, some interesting analysis can be done on how and why dependency graphs change. For example, it might be interesting to know whether a package would automatically pick a recently released version of a package for its dependency graph. Requirements can be used to determine this. Knowing whether a package would automatically pick up a new version of an existing dependency is particularly useful when we think about remediating vulnerabilities introduced by transitive dependencies.

Where does the data come from?

The main source of data is the .nuspec file of the package itself, which is available from the NuGet PackageContent API. We also use the NuGet Search API and Catalog API for version metadata fields not available in the .nuspec file.

We’d like to thank the NuGet team for helping us develop support for NuGet on Each time we develop support for a new ecosystem we discover interesting differences across ecosystems that require us to expand our backend data model and infrastructure. It was great having the official maintainers lend a hand in helping us understand the NuGet ecosystem.

What are some ways in which I can use the data?

NuGet data is available in our API, website and BigQuery dataset.

The new requirements tab available for NuGet.
The new requirements tab available for NuGet.

The data that’s available:

  • Requirements: The requirements tab has been added to the UI, the GetRequirements endpoint has been added to the API and the NuGetRequirements table has been added to the BigQuery dataset.
  • Hashes: You can use the Query API endpoint to query for the name of a mystery NuGet package by using its hash.
  • Advisories: You can check whether a NuGet package version is directly affected by an advisory, and if so, whether there is an unaffected version you could use.
  • Version metadata: Finally, there’s metadata like licenses, publish dates, descriptions, owners and reference links.

What’s next?

We’re hoping to improve our license support for NuGet, as NuGet allows multiple ways to specify a package’s license.

We’re also planning on adding requirement information for other ecosystems.

We’d love to hear what you think about our NuGet support. As mentioned earlier, NuGet is our first ecosystem to feature requirements, so we’re interested in hearing about your experiences (both good and bad) working with this new kind of data. Get in touch via email at or file an issue to our GitHub repository.

Introducing npm package SLSA provenance support

Eve Martin-Jones, Open Source Insights Team

Not knowing where your software dependencies come from leaves your codebase vulnerable to breaches, exploits and supply chain attacks. Just since the beginning of 2023 we’ve seen many examples of actors attempting to inject malicious packages into open source consumers supply chains. With SLSA provenance attestations, users can verify an artifact’s build integrity to ensure that malicious parties have not created, tampered with or replaced the code they’re running.

We are excited to announce that now shows SLSA provenance information for npm packages to provide more information and improve trust for package consumers. You can view the provenance information for supported npm packages on — for example,

Screenshot of provenance UI for the npm js-cookie package
Screenshot of provenance UI for the npm js-cookie package

SLSA provenance is metadata about how a package was built and strongly links an open source package to the build system and source code used to create it. It is part of the SLSA framework for improving supply chain security. In the provenance attestation you can find:

  • the repository and commit at which the version was built
  • details about the workflow used to create the version
  • how the version was built
  • the date the attestation was integrated into the rekor transparency log.

npm recently announced the public beta for their integration with the Sigstore project. This allows package owners to upload cryptographically verifiable SLSA provenance attestations along with their packages. In the future, integrations with the npm CLI tool will automatically verify attestations on install, making working in npm more secure by default. (However, it’s important to note that not all npm packages will have these attestations and older versions will not automatically be republished. Further information is available in npm’s announcement).

It’s great to see npm working to make the open source ecosystem more secure and we hope to see other package managers follow suit! For more information about SLSA visit the SLSA homepage. Instructions are available for npm package maintainers who want to add SLSA provenance information to their packages.

Announcing the API

Jesper Särnesjö and Nicky Ringland, Open Source Insights Team

This post was originally published on the Google Security Blog

Today, we are excited to announce the API, which provides free access to the dataset of security metadata, including dependencies, licenses, advisories, and other critical health and security signals for more than 50 million open source package versions.

Software supply chain attacks are increasingly common and harmful, with high profile incidents such as Log4Shell, Codecov, and the recent 3CX hack. The overwhelming complexity of the software ecosystem causes trouble for even the most diligent and well-resourced developers.

We hope the API will help the community make sense of complex dependency data that allows them to respond to—or even prevent—these types of attacks. By integrating this data into tools, workflows, and analyses, developers can more easily understand the risks in their software supply chains.

The power of dependency data

As part of Google’s ongoing efforts to improve open source security, the Open Source Insights team has built a reliable view of software metadata across 5 packaging ecosystems. The data set is continuously updated from a range of sources: package registries, the Open Source Vulnerability database, code hosts such as GitHub and GitLab, and the software artifacts themselves. This includes 5 million packages, more than 50 million versions, from the Go, Maven, PyPI, npm, and Cargo ecosystems—and you’d better believe we’re counting them!

We collect and aggregate this data and derive transitive dependency graphs, advisory impact reports, OpenSSF Security Scorecard information, and more. Where the website allows human exploration and examination, and the BigQuery dataset supports large-scale bulk data analysis, this new API enables programmatic, real-time access to the corpus for integration into tools, workflows, and analyses.

The API is used by a number of teams internally at Google to support the security of our own products. One of the first publicly visible uses is the GUAC integration, which uses the data to enrich SBOMs. We have more exciting integrations in the works, but we’re most excited to see what the greater open source community builds!

We see the API as being useful for tool builders, researchers, and tinkerers who want to answer questions like:

  • What versions are available for this package?
  • What are the licenses that cover this version of a package—or all the packages in my codebase?
  • How many dependencies does this package have? What are they?
  • Does the latest version of this package include changes to dependencies or licenses?
  • What versions of what packages correspond to this file?

Taken together, this information can help answer the most important overarching question: how much risk would this dependency add to my project?

The API can help surface critical security information where and when developers can act. This data can be integrated into:

  • IDE Plugins, to make dependency and security information immediately available.
  • CI/CD integrations to prevent rolling out code with vulnerability or license problems).
  • Build tools and policy engine integrations to help ensure compliance.
  • Post-release analysis tools to detect newly discovered vulnerabilities in your codebase.
  • Tools to improve inventory management and mystery file identification.
  • Visualizations to help you discover what your dependency graph actually looks like:
What you think your dependency graph looks like vs what your dependency graph actually looks like
What you think your dependency graph looks like vs what your dependency graph actually looks like

Unique features

The API has a couple of great features that aren’t available through the website.

Hash queries

A unique feature of the API is hash queries: you can look up the hash of a file’s contents and find all the package versions that contain that file. This can help figure out what version of which package you have even absent other build metadata, which is useful in areas such as SBOMs, container analysis, incident response, and forensics.

Real dependency graphs

The dependency data is not just what a package declares (its manifests, lock files, etc.), but rather a full dependency graph computed using the same algorithms as the packaging tools (Maven, npm, Pip, Go, Cargo). This gives a real set of dependencies similar to what you would get by actually installing the package, which is useful when a package changes but the developer doesn’t update the lock file. With the API, tools can assess, monitor, or visualize expected (or unexpected!) dependencies.

API in action

For a demonstration of how the API can help software supply chain security efforts, consider the questions it could answer in a situation like the Log4Shell discovery:

  • Am I affected? - A CI/CD integration powered by the free API would automatically detect that a new, critical vulnerability is affecting your codebase, and alert you to act.
  • Where? - A dependency visualization tool pulling from the API transitive dependency graphs would help you identify whether you can update one of your direct dependencies to fix the issue. If you were blocked, the tool would point you at the package(s) that are yet to be patched, so you could contribute a PR and help unblock yourself further up the tree.
  • Where else? - You could query the API with hashes of vendored JAR files to check if vulnerable log4j versions were unexpectedly hiding therein.
  • How much of the ecosystem is impacted? - Researchers, package managers, and other interested observers could use the API to understand how their ecosystem has been affected, as we did in this blog post about Log4Shell’s impact.

Getting started

The API service is globally replicated and highly available, meaning that you and your tools can depend on it being there when you need it.

It’s also free and immediately available—no need to register for an API key. It’s just a simple, unauthenticated HTTPS API that returns JSON objects:

# List the advisories affecting log4j 1.2.17
$ curl \
    | jq '.advisoryKeys[].id'

A single API call to list all the GHSA advisories affecting a specific version of log4j

Check out the API Documentation to get started, or jump straight into the code with some examples.

Dependency graph for react 15.0.0, fetched from the API and rendered using GraphViz
Dependency graph for react 15.0.0, fetched from the API and rendered using GraphViz

Securing supply chains

Software supply chain security is hard, but it’s in all our interests to make it easier. Every day, Google works hard to create a safer internet, and we’re proud to be releasing this API to help do just that, and make this data universally accessible and useful to everyone.

We look forward to seeing what you might do with the API, and would appreciate your feedback. (What works? What doesn’t? What makes it better?) You can reach us at, or by filing an issue on our GitHub repo.

Post-Advisory Exposure

James Wetter and Nicky Ringland, Open Source Insights Team

How can a user of open source software (OSS) assess their risk of exposure to a future vulnerability when taking on a new dependency?

Vulnerabilities will always find their way into software, and in an ideal world those vulnerabilities will be fixed in a reasonable amount of time. This is a critical factor for building trust between OSS maintainers and the users of their software.

This blog post looks at the events around the remediation of a vulnerability, and a few ways that trust can be established between maintainers and users of OSS. In particular we examine how often OSS packages remediate known vulnerabilities and if their users were left exposed after the vulnerability was publicly disclosed.

An ideal remediation

next-auth is an npm package that provides tools to help implement authentication for the web development framework Next.js. next-auth is popular, with almost 200,000 weekly downloads according to npm. Recently an advisory was published detailing a critical vulnerability in the next-auth package. Due to this vulnerability, an attacker could potentially gain access to another user’s account.

Fortunately for the users of next-auth, the reporter of the vulnerability and package maintainer practiced coordinated vulnerability disclosure. As a result a fixed version of next-auth was already available when this advisory was published. Both versions 4.10.3 and 3.29.10 include a patch remediating the vulnerability.

The advisory itself contains a brief timeline of key events. The vulnerability was discovered by Socket, and privately disclosed to the maintainers of next-auth on the 26th of July. The maintainers acknowledged the private disclosure within 1 hour, and released remediating versions on the 1st of August. Two days later, an advisory disclosing the vulnerability was published. The time between private disclosure and the release of a fix, the time to remediation, was approximately 5 working days.

The events surrounding a coordinated vulnerability disclosure.
The events surrounding a coordinated vulnerability disclosure.

This situation is ideal. Both the private disclosure of the vulnerability and rapid response of the package maintainers meant that the two most recent major versions both had patched versions available for users before the publication of the advisory.

By the time the advisory was published, most users of the next-auth package would be able to move to a patched version immediately with little effort. This virtually eliminated the post-advisory exposure time for the many users of the package.

What can go wrong?

Things don’t always work out as well as this, though. There are a few ways in which the process could go awry such as the discovery of a zero-day exploit, or a vulnerability in an unmaintained package.

A zero-day exploit

The events surrounding a hypothetical zero day exploit.
The events surrounding a hypothetical zero day exploit.

A zero-day exploit is when a vulnerability is being actively exploited by the time the package maintainers become aware of the issue. In these situations it may be better to publish an advisory before the maintainers have developed a patch in order to raise awareness as quickly as possible. This was the case for the well publicized remote code injection vulnerability in the popular log4j library.

In this scenario, it is not reasonable to expect the maintainers to remediate the vulnerability before the advisory is published - increased awareness is a higher priority. And as a result the users of the package will be exposed to a publicly known vulnerability until a remediation is made available, or they remove their dependency on the affected package.

An unmaintained package

The events surrounding a failed coordinated vulnerability disclosure for a hypothetical unmaintained package.
The events surrounding a failed coordinated vulnerability disclosure for a hypothetical unmaintained package.

When a vulnerability is discovered in a package that is no longer maintained there will be no response to private disclosure, leaving the reporter no choice but for the reporter to publish an advisory without a fix available.

An example of this situation is the once popular npm package parsejson. Its most recent release has an unremedied, high severity vulnerability that was publicly disclosed in 2018. But the package hasn’t seen a new release since 2016. Its GitHub repository has been archived and clearly states that it is no longer maintained. Worryingly, the package is still widely used: npm reports that the package still gets almost 250,000 weekly downloads.

It’s clear that users of OSS should not introduce new dependencies on an unmaintained package like parsejson. Existing users should remove such dependencies from their libraries and applications as quickly as they can. But it can be hard for a developer to know when one of their dependencies is no longer maintained or less actively maintained. Signals to help identify changes in the maintenance status are critical.

What usually happens after an advisory?

For our discussion here, we consider a package to have remediated an advisory when it has a release that

  1. is not affected by the advisory, and
  2. has a greater version number than all affected releases.

The semantics of versions and release differ between systems. For example PyPI uses pep440, while npm uses semantic versioning.

This definition of remediation means that if the greatest major version of a package has a fix available, the package is considered to have remediated the vulnerability even if lesser major versions remain affected. There is more to be said about packages that have multiple major versions, each of which may be fixed independently, but we will leave a discussion of the nuance of vulnerabilities and multiple major versions for another time.

Clearance rates

First let’s take a look at how many known vulnerabilities are remediated.

Across every package management system supported by, we see that most package maintainers do respond to vulnerability advisories in their packages.

There is considerable variation between ecosystems. The lower clearance rate seen in the Cargo ecosystem is expected. Within that ecosystem, there is a practice of publishing an advisory that a package is unmaintained, such as this advisory and this advisory. Such advisories are not expected to be remediated, but publishing them helps raise awareness of the package’s unmaintained state amongst its users.

Taking a closer look at individual packages, the clearance rate of vulnerabilities gives an indication of the health of the package, and consequent risk of using the package. Some packages have a very high number of known vulnerabilities in older versions, but all of the vulnerabilities have been remediated. For example

These packages are healthy and well maintained, and their high clearance rates are a good indication of that.

Post-advisory exposure time

Now let’s consider how long users are exposed to a known vulnerability without a fix. That is, the interval between the publication of an advisory and the publication of a release to remediate it. We call this the post-advisory exposure time.

The PyPI, Cargo and npm packaging systems expose the publication times for each version. Using this data we can examine the post-advisory exposure time.

The post-advisory exposure time across the PyPI, Cargo and npm ecosystems for all the vulnerabilities in the dataset.
The post-advisory exposure time across the PyPI, Cargo and npm ecosystems for all the vulnerabilities in the dataset.

At a glance these graphs paint a positive picture. Each ecosystem appears healthy, with the majority of vulnerabilities disclosed in an advisory being remediated very quickly. This demonstrates that security is a priority for most maintainers.

But it should be noted that vulnerabilities where coordinated disclosure was successful will have zero post-advisory exposure time (or even negative time!). In npm and PyPI almost 60% of the vulnerabilities in our database were remediated before the publication of the corresponding advisory. Cargo has a much lower percentage, around 16%; more on that shortly.

Let’s direct our attention to cases that did not see a coordinated vulnerability disclosure. The following histograms show the post-advisory exposure time, excluding successfully coordinated disclosures.

The post-advisory exposure time across the PyPI, Cargo and npm ecosystems for the vulnerabilities in the dataset that didn't have a coordinated disclosure.
The post-advisory exposure time across the PyPI, Cargo and npm ecosystems for the vulnerabilities in the dataset that didn't have a coordinated disclosure.

In all three systems, many vulnerabilities are remediated within 30 days of advisory publication. This includes many zero-day exploits, such as log4shell, that were fixed as quickly as possible, even without the more ideal option of a coordinated vulnerability disclosure.

In the case of Cargo, the number resolved in the first 30 days is a staggering 70% of all vulnerabilities remediated after advisory publication. This is because many maintainers choose to release the remediation on the same day the advisory is published, resulting in non-zero but very brief post-advisory exposure time.

The long tail of vulnerabilities with significant post-advisory exposure time is a valuable signal on the health of the corresponding packages. For developers taking on new dependencies, knowing that they will not be left exposed for long periods of time is critical to their security posture. For existing users of a dependency, being aware of changes to future remediation likelihood of potential vulnerabilities is equally important.

Currently it is hard to know how a given package has previously performed according to this metric. Ideally this information would be easily accessible, allowing potential and existing users to make informed decisions about their dependencies.

Mean time to remediation

The number of known vulnerabilities that a package maintainer has remediated in the past can be used to help build trust between maintainers and users of OSS. Additionally, the length of time users of a package were left exposed to known, unremedied vulnerabilities in the past can provide a more detailed characterization of a package maintainer’s response.

In addition to these signals, Mean Time to Remediation (MTTR) has been proposed as a useful indicator of the quality of a package’s maintenance.

However, the available data about advisories rarely contains timestamps for critical events in the remediation process. For example, most advisory databases, including GitHub Advisories and OSV, do not provide a timestamp field for the private disclosure of the vulnerability or the maintainers acknowledgement. And while some advisory write-ups do include an event timeline, these are quite rare.

These missing timestamps make it impossible to compute the time that elapsed between a maintainer being notified of a vulnerability and the release of a remediation, relegating MTTR to a, for now, still hypothetical metric to compute.


Vulnerabilities are an inevitable part of software development. The code reuse and efficiency gains made possible by OSS broadens the potential impact of vulnerabilities.

But cooperation between parties that discover vulnerabilities and package maintainers reduces the time that users are left exposed to publicly known vulnerabilities. Thanks to the hard work of OSS maintainers, there is no post-advisory exposure for the majority of vulnerabilities in our advisory database.

Developers should still prepare for less ideal outcomes. Every dependency they introduce increases the risk of exposure to future vulnerabilities. The clearance rate and post-advisory exposure time for past advisories can provide users of OSS assurance about the quality of maintenance their dependencies receive. While past performance may not always predict future behavior, it can be used as a valuable signal to help make informed decisions.

After the Advisory

Nicky Ringland and James Wetter, Open Source Insights Team

This blog is based on a presentation given by Nicky Ringland at Google Open Source Live.

Open source software powers the world. Open source libraries allow developers to build things faster, organizations to be more nimble, and all of us to be more productive.

But dependencies bring complexity. Popular open source packages are often used directly or indirectly by a significant portion of the packages within an ecosystem. As a result, a vulnerability in a popular package can have a massive impact across an entire ecosystem.

Different software ecosystems have different conventions for specifying dependency requirements and different algorithms for resolving them. We will take a look at a couple of large profile incidents that discuss some of these differences.

The amplification of vulnerability impact

To measure the potential impact of a vulnerability, we can look at how many dependents it has. That is, how many other packages that use a specific version that is affected by a vulnerability. We can get a view of an ecosystem by looking at all package versions that are affected - either directly or indirectly - by a vulnerability.

First off: packages that are directly affected. At the time of writing, across all the packaging systems supported by, over 200 thousand package versions (0.4%) are directly named as vulnerable by a known advisory.

In contrast, almost 15 million package versions (33%) are affected only indirectly, by having an affected package in their dependency graph. That’s two orders of magnitude difference!

Dependencies magnify the impact of vulnerabilities.
Dependencies magnify the impact of vulnerabilities.

That underpins just how hard it is to fix a vulnerability in an ecosystem. When a package explicitly named by an advisory publishes a fix for the issue, the story is far from over. Many users of the packaging ecosystem will still be at risk, because they depend on vulnerable versions of the package deep within their dependency graphs. Fixing the directly affected package is often only the tip of the iceberg.

Addressing vulnerabilities in your dependencies

There are several ways an application maintainer could mitigate a vulnerability affecting one of their dependencies. Let’s be kind to our hypothetical maintainer and consider a simple dependency tree with two layers of dependencies.

An application may have direct and indirect dependencies.
An application may have direct and indirect dependencies.

If this maintainer is lucky, they depend on the affected package directly. That means as soon as the affected package publishes a fixed version they can update their project or application to depend on the fixed version.

A vulnerability in a direct dependency.
A vulnerability in a direct dependency.

But if the vulnerable package is among their indirect dependencies the situation could be much more complex.

In the best case scenario, the intermediate packages already depend on the patched version.

A vulnerability in an indirect dependency.
A vulnerability in an indirect dependency.

If this is not the case, our hypothetical maintainer may still have a course of action. To update to the fixed version of the indirect dependency the maintainer may be able to specify the fixed version as a minimum for the entire dependency graph. For this to work, however, the fixed version of the affected package and its direct dependents must be compatible. If not, the maintainer may have to wait for a new release of the intermediate dependent.

An indirect dependency can often have its version pinned, but incompatibile intermediate dependencies may be a problem.
An indirect dependency can often have its version pinned, but incompatibile intermediate dependencies may be a problem.

Another alternative is to remove the dependency on the affected package. But this often involves considerable effort; you would never have added a dependency without good reason, right?

A vulnerability in a dependency can be mitigated by removing that dependency.
A vulnerability in a dependency can be mitigated by removing that dependency.

In practice, dependency trees are rarely so simple and clean. Usually they are complex, interconnected graphs. Just take a look at the dependency graphs for popular frameworks and tools like express or kubernetes.

The dependency graph of a version of express.
The dependency graph of a version of express.

These complex graphs can make remediating a vulnerability far more difficult than the simple examples given above. There may be many paths through which a fix must propagate before it gets to you. Or, in order to remove a dependency, you might need to remove a significant portion of your dependency graph.

For example, consider the many paths by which one package depends on a vulnerable version of log4j:

Various paths from through a dependency graph to a vulnerable version of log4j.
Various paths from through a dependency graph to a vulnerable version of log4j.

With this in mind, perhaps you can imagine why it often takes a long time for a patched version of a popular package to roll out to the ecosystem.

log4shell in the Maven Central ecosystem

On December 9th last year, over 17,000 of the Java packages available from Maven central were impacted by the log4j vulnerabilities, known as log4shell, resulting in widespread fallout across the software industry. The vulnerabilities allowed an attacker to perform remote code execution by exploiting the insecure JNDI lookups feature exposed by the logging library log4j. This exploitable feature was present in multiple versions, and was enabled by default in many versions of the library. We wrote about this incident shortly after it occurred in a previous blog.

A new version of log4j with the vulnerability patched (albeit with few false starts due to incomplete fixes) was available almost immediately. So once that patched version was published had the ecosystem freed itself of log4shell? Unfortunately not. Part of what makes fixing log4shell hard is Java’s conventions on how dependency requirements are specified, and Maven’s dependency resolution algorithm itself.

In the Java ecosystem, it’s common practice to specify “soft” version requirements. That is, the dependent package specifies a single version for each dependency, which is usually the version that ends up being used. (The dependency resolution algorithm may choose a different version under certain rare conditions – for example, a different version already in the graph). While it is possible to specify ranges of suitable versions, this is unusual. More than 99% of dependency requirements in the Maven Central ecosystem are specified using soft requirements.

Here’s where Maven’s dependency resolution algorithm comes in. Since almost all the time, a specific version has been specified, that’s almost always the version that the dependency resolution will pick. So if a newer version with that important new bug fix is released, it won’t be included automatically. It usually requires explicit action by the maintainer to update the dependency requirements to a patched version.

As a result of the pervasive use of 'soft' dependency requirements old, vulnerable versions of a package may continue to be depended on long after the release of a patched version.
As a result of the pervasive use of 'soft' dependency requirements old, vulnerable versions of a package may continue to be depended on long after the release of a patched version.

In this case, consumers of any one of the 17,000 odd packages affected by the log4j vulnerabilities would likely still depend on an affected version of log4j, even after the first fix was published. Ideally the maintainers of around 4,000 packages that directly depend on log4j would promptly release a new version of their package that explicitly requires a fixed version of log4j. Then the maintainers of packages that depend on those packages can update their version requirements, and then maintainers of those packages, and so on. There are methods to pin the version of indirect dependencies accelerating this process, but many consumers rely on the default behavior of their tools.

It’s been over six months since the log4 advisory was disclosed. How well has the underlying fix to log4shell propagated throughout the ecosystem? A little less than a week after the disclosure around 13% of affected packages had remediated the issue by releasing a new version. 10 days after disclosure this number had risen to around 25%. Now a few months after that we see around 40% of the affected packages have remediated the problem. Considering how widespread the problem was, and the complexity of the dependencies between packages, this is amazing progress, but there’s clearly a lot more to go.

Default versions: new or old?

Package managers differ in which versions they choose to install by default. For example, systems like Maven or Go err on the side of choosing earlier matching versions, while npm and Pip tend to choose later versions. This design choice can have a big impact on how a fix rolls out or, conversely, how quickly an exploit can propagate.

Choosing the earlier versions has the benefit of stability; dependency graphs remain stable whether you install today or tomorrow, even if new versions are released. The downside is that the consumer must be conscientious in updating their dependencies when security issues arise.

Choosing the later versions has the benefit of currency; you get the latest fixes automatically just by reinstalling. The downside here is that your dependencies can change underfoot, sometimes in dramatic and unexpected ways.

With this in mind, if log4shell had occurred in the npm or PyPI ecosystems the story would have been quite different. In these ecosystems, packages typically ask for the most recent compatible versions of their dependencies.

The conventions of npm mean fixes can propagate throughout the ecosystem automatically.
The conventions of npm mean fixes can propagate throughout the ecosystem automatically.

Looking at the dependency requirements across all versions of all packages in npm we find around three quarters use the caret (^) or tilde (~) allowing a new patch or minor version of the dependency to be automatically selected when available. When adhering to semantic versioning, this means that many users will use the newest release with a compatible API by default.

This practice would likely have been a substantial benefit in remediating a log4shell-like event, where a vulnerability is discovered in widely used versions of a popular package.

But as we shall see, sometimes we really, really don’t want to use the latest version.

The case of colors

In early January 2022, the developer of the popular npm packages colors and faker intentionally published several releases containing breaking changes. These were picked up rapidly due to the npm resolution algorithm preferencing recent releases, and the norm in javascript of using dependency requirements that allow the use of new compatible versions automatically.

The conventions of npm mean new malicious code can propagate throughout the ecosystem rapidly.
The conventions of npm mean new malicious code can propagate throughout the ecosystem rapidly.

At the time of the incident, more than 100,000 packages’ most recent releases depended on a version of colors, and around half of them had a dependency on a problematic version. The following graph shows the dependency flow in the ecosystem over the 72 hours where the action happened.

The number of packages dependent on different versions of colors.
The number of packages dependent on different versions of colors.

About half the packages depending on colors remained unaffected throughout the incident because they depended on earlier versions of colors. But the other half of packages had some rapid and widespread changes in the exact version of colors that would have been used depending on the time at which their dependencies were resolved.

The first problematic version was 1.4.44-liberty-2. Due to version naming conventions this isn’t considered a stable version and as a result it wasn’t depended on by many packages.

A few hours later version 1.4.1 was released, and almost all packages using the 1.4 minor version immediately began to depend on this problematic version. Several hours later, 1.4.2 was released, and again most packages affected by the incident immediately depended on this new problematic version. After a few more hours npm stepped in and removed all the bad versions of colors, at which point all dependents moved back to safe versions.

The speed of this incident impacting the ecosystem was rapid. But so too was the rapid response of maintainers. Between the initial release of bad versions and their removal from npm, a period of less than 72 hours, nearly half of all affected packages were able to mitigate the issue. A small number of packages were able to remove their dependency on colors, about 4% of affected packages, seen as a drop in total number of dependent packages. Many more packages, 40% of those affected, were able to pin the version of colors being used to a safe version. This can be seen in the gradual increase of packages depending on an unaffected 1.4.x version.

Interestingly this rapid mitigation was the work of very few people. Just a little over 1% of the affected packages actually made a release during this time period. But their work resulted in 43% of the total affected packages mitigating the issue. This is a result of the same use of open dependency requirements that allowed the rapid spread of the issue and enabled rapid mitigation.

Every dependency is a trust relationship

The colors and log4shell incidents were very similar in terms of wide-reaching impact, but quite different in onset and response. In the case of log4shell, a new vulnerability was discovered in old and widely used versions, resulting in a need for dependents to move to a new release of the package. In the case of colors, a new release introduced breaking changes. This resulted in an initial automated surge to the problematic version, followed by a concerted effort for dependents to move to an older release of the package.

While the widespread use of open dependency constraints in npm led to a rapid and widespread impact of colors, it was also helpful in its mitigation. Conversely Maven’s approach of favoring stability resulted in difficulty resolving log4shell, but also means Maven is much less susceptible to a colors-type incident. Neither approach is obviously superior, just different.

While there is no silver bullet solution, there are best practices that consumers, maintainers, and packaging system developers can observe to reduce risk. Always understand your dependencies and why they were chosen, and always make sure your dependency requirements are well maintained.

Understanding the Impact of Apache Log4j Vulnerability

James Wetter and Nicky Ringland, Open Source Insights Team

More than 17,000 Java packages, amounting to over 4% of the Maven Central repository (the most significant Java package repository), have been impacted by the recently disclosed log4j vulnerabilities (1, 2), with widespread fallout across the software industry.1 The vulnerabilities allow an attacker to perform remote code execution by exploiting the insecure JNDI lookups feature exposed by the logging library log4j. This exploitable feature was enabled by default in many versions of the library.

This vulnerability has captivated the information security ecosystem since its disclosure on December 9th because of both its severity and widespread impact. As a popular logging tool, log4j is used by tens of thousands of software packages (known as artifacts in the Java ecosystem) and projects across the software industry. User’s lack of visibility into their dependencies and transitive dependencies has made patching difficult; it has also made it difficult to determine the full blast radius of this vulnerability. Using Open Source Insights, a project to help understand open source dependencies, we surveyed all versions of all artifacts in the Maven Central Repository to determine the scope of the issue in the open source ecosystem of JVM based languages, and to track the ongoing efforts to mitigate the affected packages.

How widespread is the log4j vulnerability?

As of December 16, 2021, we found that over 17,000 of the available Java artifacts from Maven Central depend on the affected log4j code. This means that more than 4% of all packages on Maven Central have at least one version that is impacted by this vulnerability.1 (These numbers do not encompass all Java packages, such as directly distributed binaries, but Maven Central is a strong proxy for the state of the ecosystem.)

As far as ecosystem impact goes, 4% is enormous. The average ecosystem impact of advisories affecting Maven Central is 2%, with the median less than 0.1%.

Maven ecosystem affected by the vulnerability
Maven ecosystem affected by the vulnerability

Direct dependencies account for around 3,500 of the affected artifacts, meaning that any of its versions depend upon an affected version of log4j-core as described in the CVEs. The majority of affected artifacts come from indirect dependencies (that is, the dependencies of one’s own dependencies), meaning log4j is not explicitly defined as a dependency of the artifact, but gets pulled in as a transitive dependency.

Direct dependencies vs. indirect dependencies
Direct dependencies vs. indirect dependencies

What is the current progress in fixing the open source JVM ecosystem?

We counted an artifact as fixed if the artifact had at least one version affected and has released a greater stable version (according to semantic versioning) that is unaffected. An artifact affected by log4j is considered fixed if it has updated to 2.16.0 or removed its dependency on log4j altogether.

At the time of writing, nearly five thousand of the affected artifacts have been fixed. This represents a rapid response and mammoth effort both by the log4j maintainers and the wider community of open source consumers. That leaves over 12,000 artifacts affected, many of which are dependent on another artifact to patch (the transitive dependency) and are likely blocked.

An example of affected artifact blocked by an intermediate dependency
An example of affected artifact blocked by an intermediate dependency

Why is fixing the JVM ecosystem hard?

Most artifacts that depend on log4j do so indirectly. The deeper the vulnerability is in a dependency chain, the more steps that may be required for it to be fixed. The following diagram shows a histogram of how deeply an affected log4j package (core or api) first appears in consumers dependency graphs. For greater than 80% of the packages, the vulnerability is more than one level deep, with a majority affected five levels down (and some as many as nine levels down). These packages may require fixes throughout all parts of the tree, starting from the deepest dependencies first.

Depth of log4j dependencies
Depth of log4j dependencies

Another difficulty is caused by ecosystem-level choices in the dependency resolution algorithm and requirement specification conventions.

In the Java ecosystem, it’s common practice to specify “soft” version requirements — exact versions that are used by the resolution algorithm if no other version of the same package appears earlier in the dependency graph. Propagating a fix often requires explicit action by the maintainers to update the dependency requirements to a patched version.

This practice is in contrast to other ecosystems, such as npm, where it’s common for developers to specify open ranges for dependency requirements. Open ranges allow the resolution algorithm to select the most recently released version that satisfies dependency requirements, thereby pulling in new fixes. Consumers can get a patched version on the next build after the patch is available, which propagates up the dependencies quickly. (This approach is not without its drawbacks; pulling in new fixes can also pull in new problems.)

How long will it take for this vulnerability to be fixed across the entire ecosystem?

It’s hard to say. We looked at all publicly disclosed critical advisories affecting Maven packages to get a sense of how quickly other vulnerabilities have been fully addressed. Less than half (48%) of the artifacts affected by a vulnerability have been fixed, so we might be in for a long wait, likely years.

But things are looking promising on the log4j front. After less than a week, around 25% of affected artifacts have been fixed. This, more than any other stat, speaks to the massive effort by open source maintainers, information security teams and consumers across the globe.

Where to focus next?

Thanks and congratulations are due to the open source maintainers and consumers who have already upgraded their versions of log4j. As part of our investigation, we pulled together a list of 500 affected packages with some of the highest transitive usage. If you are a maintainer or user helping with the patching effort, prioritizing these packages could maximize your impact and unblock more of the community.

We encourage the open source community to continue to strengthen security in these packages by enabling automated dependency updates and adding security mitigations. Improvements such as these could qualify for financial rewards from the Secure Open Source Rewards program.

You can explore your package dependencies and their vulnerabilities by using Open Source Insights.

  1. When this blog post was initially published, count numbers included all packages dependent on either log4j-core or log4j-api, as both were listed as affected in the CVE. The numbers have been updated to account for only packages dependent on log4j-core.  2

Introducing PyPI Support for

Paul Mathews and the Open Source Insights Team

We’re pleased to announce now has support for Python packages hosted on the Python Package Index (PyPI). That means we have over 300k—and counting—Python packages for your perusal, from boto3 to pandas.

Where does the data come from?

We use PyPI’s RSS Feeds to stay abreast of new and updated packages, with an occasional full sync from the Simple Repository API. For each package version, we fetch metadata from the JSON API and analyze it to resolve its dependencies, determine the license, and so on.

Dependency resolution is complex in any language, and Python is no exception. Sometimes you might see an error message about a particular version of a package. The most common reason for this is packages that only provide a source distribution that specifies the dependencies in a—which is hard to run safely and may not even be deterministic. This is not a problem with wheels as they do not require executing arbitrary Python code to understand the dependencies. Of course there are any number of other things that can go wrong, and Python has a long history of packaging formats, so if you find anything not working as expected, don’t hesitate to get in touch.

Where do the dependencies come from?

We periodically resolve the full dependencies of every package version we know about. In pip terms, the graph we show for version 1.0.0 of package a consists of the packages that would be installed by running pip install a==1.0.0 in a clean environment with recent versions of setuptools and wheel available.

These graphs are dependent on the versions of both Python and pip, as well as the operating system, CPU architecture, and so on. It’s not uncommon for packages to publish different wheels for various different combinations of all of these, and for each release to have its own metadata with potentially distinct dependencies. Currently we perform resolution as if we were runnning pip 21.1.3 with Python 3.9 on an x86_64 manylinux compatible platform, with more combinations on the way. We think it’s an accurate reproduction but if you see anything unexpected, please let us know!

What’s next

We’re excited to add PyPI to our set of supported language ecosystems, and epecially keen to start digging into the data and do some comparative analysis. From our first look, there are plenty of interesting things to uncover, for instance:

  • 4 of the 5 most depended on packages are all dependencies of the 6th most depended on package: requests
  • more than half of all package versions on PyPI have zero dependencies, compared to ≈15-25% across Go, npm, Cargo and Maven
  • this small package has one of the lowest ratios of direct to indirect dependents we’ve seen across all package ecosystems.

We’re also working on improving our license recognition and figuring out how to show the differences enabling various extras makes to the dependency graphs.

So slither on in and start exploring! We’ll keep digging into the data and keep you posted on what we discover!

Introducing Open Source Insights

Nicky Ringland and the Open Source Insights team

Modern software is more than just some lines of code checked into a repository. To build almost any program, one must also install packages from other developers. These external dependencies are critical components of today’s software environment, and tooling has been created to make it easy to install dependencies and update them as required. As a result, the past few years have witnessed a phenomenal growth in the open source ecosystem as well as a marked increase in the average number of dependencies for a given package. Meanwhile, many of these packages are being changed—fixed, expanded, updated—regularly.

The rate of change is significant. Our analysis shows that roughly 15% of the packages in npm see changes to their dependency sets each day, while a majority of the change is in packages that are widely used.

This activity affects not just your own software, and not even just the software you call upon, but the entire set of your software’s dependencies, which may be much larger than those listed explicitly by your project. It is common to see one package use a handful of other packages that in turn have a hundred or more dependencies of their own. Many of the most commonly used packages in open source have large dependency trees that will be pulled in by the installation process.

Today’s software is therefore built upon a constantly-changing foundation, and keeping track of that churn is challenging. Your package changes, your dependencies change, their dependencies change, and so on. Even the most diligent developers struggle to keep up beyond letting the tooling download updates to all the dependencies from time to time. Tooling helps manage the updates, but cannot guarantee what the right update is, or when the time is right to apply it.

It’s easy to miss important problems deep in the dependencies, such as security vulnerabilities, license conflicts, or other issues. The tools just do what they are told, and if a nested dependency has an issue, it will be installed regardless. Systems have been compromised or exploited by dependencies that acquired malicious changes that were undetected, sometimes for long periods.

The Open Source Insights project aims to help. It collects information about open source projects—source code, licenses, releases, vulnerabilities, owners, and more—and gathers it into a single location, making it accessible. These interfaces help developers and project owners see the full dependency graph of their projects and can use it to track release activity, vulnerabilities, and other information such as licenses that are used by the components, regardless of how deeply they are nested inside the dependencies.

In short, all the information about a package is connected to all the other packages that depend upon it, and Insights shows the connections. For instance, if your code depends on a package that has a security vulnerability, even if that vulnerability is in a package 10 dependency hops away in a package that you don’t even know about, the Insights page for your package will tell you about it.

The Insights project also helps developers see the importance of their project by showing the projects that depend on them—their dependents. Even a small project is important if a large number of other projects depend on it, either directly or through transitive dependencies.

This blog

To build we dove deep into the fundamentals of several different package management systems, collecting and organizing the metadata of millions of packages, and implementing our own bug-for-bug compatible semver parsers, constraint matchers, and dependency resolvers.

Along the way, we’ve learnt about the wider problem space, and the varied challenges that await the unsuspecting programmer. The tools are changing, and inconsistent, and often poorly understood. Many package management systems were not designed with today’s security concerns in mind. Semver constraints are not formally specified, and are implemented arbitrarily by different package managers. There is not widespread agreement on foundational questions such as whether it is better to select the newest or oldest matching version, or whether the ability to “pin” versions is a good or bad thing (and what that even means). Also whether it is good or bad (or possible or impossible) to be able to include multiple versions of the same package in a given program.

In future articles we will explore these issues in detail, comparing the approaches of various communities, so as to contribute to the conversations that push the open source ecosystem forward. We hope that we can converge, as an industry, on some fundamental “good ideas” in the space of managing software dependencies.