Skip to content

Audit Rules🔗

This page documents each of the audits currently implemented in zizmor.

See each audit's section for its scope, behavior, and other information.

Legend:

Type Examples Introduced in Works offline Enabled by default
The kind of audit ("Workflow" or "Action") Links to vulnerable examples Added to zizmor in this version The audit works with --offline The audit needs to be explicitly enabled with --pedantic

artipacked🔗

Type Examples Introduced in Works offline Enabled by default
Workflow artipacked.yml v0.1.0 ✅ ✅

Detects local filesystem git credential storage on GitHub Actions, as well as potential avenues for unintentional persistence of credentials in artifacts.

By default, using actions/checkout causes a credential to be persisted in the checked-out repo's .git/config, so that subsequent git operations can be authenticated.

Subsequent steps may accidentally publicly persist .git/config, e.g. by including it in a publicly accessible artifact via actions/upload-artifact.

However, even without this, persisting the credential in the .git/config is non-ideal unless actually needed.

Other resources:

Remediation🔗

Unless needed for git operations, actions/checkout should be used with persist-credentials: false.

If the persisted credential is needed, it should be made explicit with persist-credentials: true.

artipacked.yml
on: push

jobs:
  artipacked:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
artipacked.yml
on: push

jobs:
  artipacked:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          persist-credentials: false

dangerous-triggers🔗

Type Examples Introduced in Works offline Enabled by default
Workflow pull-request-target.yml v0.1.0 ✅ ✅

Detects fundamentally dangerous GitHub Actions workflow triggers.

Many of GitHub's workflow triggers are difficult to use securely. This audit checks for some of the biggest offenders:

  • pull_request_target
  • workflow_run

These triggers are dangerous because they run in the context of the target repository rather than the fork repository, while also being typically triggerable by the latter. This can lead to attacker controlled code execution or unexpected action runs with context controlled by a malicious fork.

Other resources:

Remediation🔗

The use of dangerous triggers can be difficult to remediate, since they don't always have an immediate replacement.

Replacing a dangerous trigger with a safer one (or keeping the dangerous trigger, but eliminating the risk of code execution) requires case-by-case consideration.

Some general pointers:

  • Replace workflow_run triggers with workflow_call: this will require re-tooling the workflow to be a reusable workflow.
  • Replace pull_request_target with pull_request, unless you absolutely need repository write permissions (e.g. to leave a comment or make other changes to the upstream repo).
  • Never run PR-controlled code in the context of a pull_request_target-triggered workflow.
  • Avoid attacker-controllable flows into GITHUB_ENV in both workflow_run and pull_request_target workflows, since these can lead to arbitrary code execution.

excessive-permissions🔗

Type Examples Introduced in Works offline Enabled by default
Workflow excessive-permissions.yml v0.1.0 ✅ ✅

Detects excessive permissions in workflows, both at the workflow level and individual job levels.

Users frequently over-scope their workflow and job permissions, or set broad workflow-level permissions without realizing that all jobs inherit those permissions.

Furthermore, users often don't realize that the default GITHUB_TOKEN permissions can be very broad, meaning that workflows that don't configure any permissions at all can still provide excessive credentials to their individual jobs.

Remediation🔗

In general, permissions should be declared as minimally as possible, and as close to their usage site as possible.

In practice, this means that workflows should almost always set permissions: {} at the workflow level to disable all permissions by default, and then set specific job-level permissions as needed.

For example:

excessive-permissions.yml
on:
  release:
    types:
      - published

name: release

permissions:
  id-token: write # trusted publishing + attestations

jobs:
  build:
    name: Build distributions 📦
    runs-on: ubuntu-latest
    steps:
      - # omitted for brevity

  publish:
    name: Publish Python 🐍 distributions 📦 to PyPI
    runs-on: ubuntu-latest
    needs: [build]

    steps:
      - name: Download distributions
        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4
        with:
          name: distributions
          path: dist/

      - name: publish
        uses: pypa/gh-action-pypi-publish@release/v1
excessive-permissions.yml
on:
  release:
    types:
      - published

name: release

permissions: {}

jobs:
  build:
    name: Build distributions 📦
    runs-on: ubuntu-latest
    steps:
      - # omitted for brevity

  publish:
    name: Publish Python 🐍 distributions 📦 to PyPI
    runs-on: ubuntu-latest
    needs: [build]
    permissions:
      id-token: write # trusted publishing + attestations

    steps:
      - name: Download distributions
        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4
        with:
          name: distributions
          path: dist/

      - name: publish
        uses: pypa/gh-action-pypi-publish@release/v1

hardcoded-container-credentials🔗

Type Examples Introduced in Works offline Enabled by default
Workflow hardcoded-credentials.yml v0.1.0 ✅ ✅

Detects Docker credentials (usernames and passwords) hardcoded in various places within workflows.

Remediation🔗

Use encrypted secrets instead of hardcoded credentials.

hardcoded-container-credentials.yml
on:
  push:

jobs:
  test:
    runs-on: ubuntu-latest
    container:
      image: fake.example.com/example
      credentials:
        username: user
        password: hackme
    services:
      service-1:
        image: fake.example.com/anotherexample
        credentials:
          username: user
          password: hackme
    steps:
      - run: echo 'hello!'
hardcoded-container-credentials.yml
on:
  push:

jobs:
  test:
    runs-on: ubuntu-latest
    container:
      image: fake.example.com/example
      credentials:
        username: user
        password: ${{ secrets.REGISTRY_PASSWORD }}
    services:
      service-1:
        image: fake.example.com/anotherexample
        credentials:
          username: user
          password: ${{ secrets.REGISTRY_PASSWORD }} # (1)!
    steps:
      - run: echo 'hello!'
  1. This may or may not be the same credential as above, depending on your configuration.

impostor-commit🔗

Type Examples Introduced in Works offline Enabled by default
Workflow, Action impostor-commit.yml v0.1.0 ❌ ✅

Detects commits within a repository action's network that are not present on the repository itself, also known as "impostor" commits.

GitHub represents a repository and its forks as a "network" of commits. This results in ambiguity about where a commit comes from: a commit that exists only in a fork can be referenced via its parent's owner/repo slug, and vice versa.

GitHub's network-of-forks design can be used to obscure a commit's true origin in a fully-pinned uses: workflow reference. This can be used by an attacker to surreptitiously introduce a backdoored action into a victim's workflows(s).

A notable historical example of this is github/dmca@565ece4, which appears to be on github/dmca is but really on a fork (with an impersonated commit author).

Other resources:

Remediation🔗

Impostor commits are visually indistinguishable from normal best-practice hash-pinned actions.

Always carefully review external PRs that add or change hash-pinned actions by consulting the claimant repository and confirming that the commit actually exists within it.

The only remediation, once discovered, is to replace the impostor commit within an authentic commit (or an authentic tag/branch reference).

known-vulnerable-actions🔗

Type Examples Introduced in Works offline Enabled by default
Workflow, Action known-vulnerable-actions.yml v0.1.0 ❌ ✅

Detects actions with known, publicly disclosed vulnerabilities that are tracked in the GitHub Advisories database. Examples of commonly disclosed vulnerabilities in GitHub Actions include credential disclosure and code injection via template injection.

Remediation🔗

If the vulnerability is applicable to your use: upgrade to a fixed version of the action if one is available, or remove the action's usage entirely.

ref-confusion🔗

Type Examples Introduced in Works offline Enabled by default
Workflow, Action ref-confusion.yml v0.1.0 ❌ ✅

Detects actions that are pinned to confusable symbolic refs (i.e. branches or tags).

Like with impostor commits, actions that are used with a symbolic ref in their uses: are subject to a degree of ambiguity: a ref like @v1 might refer to either a branch or tag ref.

An attacker can exploit this ambiguity to publish a branch or tag ref that takes precedence over a legitimate one, delivering a malicious action to pre-existing consumers of that action without having to modify those consumers.

Remediation🔗

Switch to hash-pinned actions.

self-hosted-runner🔗

Type Examples Introduced in Works offline Enabled by default
Workflow self-hosted.yml v0.1.0 ✅ ❌

Note

This is a --pedantic only audit, due to zizmor's limited ability to analyze runner configurations themselves. See #34 for more details.

Detects self-hosted runner usage within workflows.

GitHub supports self-hosted runners, which behave similarly to GitHub-hosted runners but use client-managed compute resources.

Self-hosted runners are very hard to secure by default, which is why GitHub does not recommend their use in public repositories.

Other resources:

Remediation🔗

In general, self-hosted runners should only be used on private repositories. Exposing self-hosted runners to potential public use is always a security risk.

In practice, there are many cases (such as custom host configurations) where a self-hosted runner is needed on a public repository. In these cases, there are steps you can take to minimize their risk:

  1. Require manual approval on workflows for all external contributors. This can be configured at repository, workflow, or enterprise-wide levels. See GitHub's docs for more information.
  2. Use only ephemeral ("just-in-time") runners. These runners are created just-in-time to perform one job and are destroyed immediately afterwards, making it harder (but not impossible) for an attacker to maintain persistence.

template-injection🔗

Type Examples Introduced in Works offline Enabled by default
Workflow, Action template-injection.yml v0.1.0 ✅ ✅

Detects potential sources of code injection via template expansion.

GitHub Actions allows workflows to define template expansions, which occur within special ${{ ... }} delimiters. These expansions happen before workflow and job execution, meaning the expansion of a given expression appears verbatim in whatever context it was performed in.

Template expansions aren't syntax-aware, meaning that they can result in unintended shell injection vectors. This is especially true when they're used with attacker-controllable expression contexts, such as github.event.issue.title (which the attacker can fully control by supplying a new issue title).

Other resources:

Remediation🔗

The most common forms of template injection are in run: and similar code-execution blocks. In these cases, an inline template expansion can typically be replaced by an environment variable whose value comes from the expanded template.

This avoids the vulnerability, since variable expansion is subject to normal shell quoting/expansion rules.

Tip

To fully remediate the vulnerability, you should not use ${{ env.VARNAME }}, since that is still a template expansion. Instead, you should use ${VARNAME} to ensure that the shell itself performs the variable expansion.

Tip

When switching to ${VARNAME}, keep in mind that different shells have different environment variable syntaxes. In particular, Powershell (the default shell on Windows runners) uses ${env:VARNAME}.

To avoid having to specialize your handling for different runners, you can set shell: sh or shell: bash.

template-injection.yml
- name: Check title
  run: |
    title="${{ github.event.issue.title }}"
    if [[ ! $title =~ ^.*:\ .*$ ]]; then
      echo "Bad issue title"
      exit 1
    fi
template-injection.yml
- name: Check title
  run: |
    title="${ISSUE_TITLE}"
    if [[ ! $title =~ ^.*:\ .*$ ]]; then
      echo "Bad issue title"
      exit 1
    fi
  env:
    ISSUE_TITLE: ${{ github.event.issue.title }}

use-trusted-publishing🔗

Type Examples Introduced in Works offline Enabled by default
Workflow pypi-manual-credential.yml v0.1.0 ✅ ✅

Detects packaging workflows that could use Trusted Publishing.

Some packaging ecosystems/indices (like PyPI and RubyGems) support "Trusted Publishing," which is an OIDC-based "tokenless" authentication mechanism for uploading to the index from within a CI/CD workflow.

This "tokenless" flow has significant security benefits over a traditional manually configured API token, and should be preferred wherever supported and possible.

Other resources:

Remediation🔗

In general, enabling Trusted Publishing requires a one-time change to your package's configuration on its associated index (e.g. PyPI or RubyGems).

Once your Trusted Publisher is registered, see pypa/gh-action-pypi-publish or rubygems/release-gem for canonical examples of using it.

unpinned-uses🔗

Type Examples Introduced in Works offline Enabled by default
Workflow, Action unpinned.yml v0.4.0 ✅ ✅

Detects "unpinned" uses: clauses.

When a uses: clause is not pinned by branch, tag, or SHA reference, GitHub Actions will use the latest commit on the referenced repository (or, in the case of Docker actions, the :latest tag).

This can represent a (small) security risk, as it leaves the calling workflow at the mercy of the callee action's default branch.

When used with --pedantic, this audit will also flag pinned-but-unhashed uses:. For example, actions/checkout@v4 will not be flagged by default, but would be flagged with --pedantic.

Remediation🔗

For repository actions (like actions/checkout): add a branch, tag, or SHA reference.

For Docker actions (like docker://ubuntu): add an appropriate :{version} suffix.

A before/after example is shown below.

unpinned-uses.yml
name: unpinned-uses
on: [push]

jobs:
unpinned-uses:
    runs-on: ubuntu-latest
    steps:
    - uses: actions/checkout
      with:
      persist-credentials: false

    - uses: docker://ubuntu
      with:
      entrypoint: /bin/echo
      args: hello!
unpinned-uses.yml
name: unpinned-uses
on: [push]

jobs:
unpinned-uses:
    runs-on: ubuntu-latest
    steps:
    - uses: actions/checkout@v4 # (1)!
      with:
      persist-credentials: false

    - uses: docker://ubuntu:24.04
      with:
      entrypoint: /bin/echo
      args: hello!
  1. Or actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 for a SHA-pinned action.

insecure-commands🔗

Type Examples Introduced in Works offline Enabled by default
Workflow, Action insecure-commands.yml v0.5.0 ✅ ✅

Detects opt-in for executing insecure workflow commands.

Workflow commands (like ::set-env and ::add-path) were deprecated by GitHub in 2020 due to their inherent weaknesses (e.g., allowing any command with the ability to emit to stdout to inject environment variables and therefore obtain code execution).

However, users can explicitly re-enable them by setting the ACTIONS_ALLOW_UNSECURE_COMMANDS environment variable at the workflow, job, or step level.

Other resources:

Remediation🔗

In general, users should use for GitHub Actions environment files (like GITHUB_PATH and GITHUB_OUTPUT) instead of using workflow commands.

insecure-commands
- name: Setup my-bin
  run: |
    echo "::add-path::$HOME/.local/my-bin"
  env:
    ACTIONS_ALLOW_UNSECURE_COMMANDS: true
insecure-commands
- name: Setup my-bin
  run: |
    echo "$HOME/.local/my-bin" >> "$GITHUB_PATH"

github-env🔗

Type Examples Introduced in Works offline Enabled by default
Workflow, Action github-env.yml v0.6.0 ✅ ✅

Detects dangerous writes to the GITHUB_ENV and GITHUB_PATH environment variables.

When used in workflows with dangerous triggers (such as pull_request_target and workflow_run), GITHUB_ENV and GITHUB_PATH can be an arbitrary code execution risk:

  • If the attacker is able to set arbitrary variables or variable contents via GITHUB_ENV, they may be able to set LD_PRELOAD or otherwise induce code execution implicitly within subsequent steps.
  • If the attacker is able to add an arbitrary directory to the $PATH via GITHUB_PATH, they may be able to execute arbitrary code by shadowing ordinary system executables (such as ssh).

Other resources:

Remediation🔗

In general, you should avoid modifying GITHUB_ENV and GITHUB_PATH within sensitive workflows that are attacker-triggered, like pull_request_target.

If you absolutely must use GITHUB_ENV or GITHUB_PATH, avoid passing attacker-controlled values into either. Stick with literal strings and values computed solely from trusted sources.

If you need to pass state between steps, consider using GITHUB_OUTPUT instead.

cache-poisoning🔗

Type Examples Introduced in Works offline Enabled by default
Workflow cache-poisoning.yml v0.10.0 ✅ ✅

Detects potential cache-poisoning scenarios in release workflows.

Caching and restoring build state is a process eased by utilities provided by GitHub, in particular actions/cache and its "save" and "restore" sub-actions. In addition, many of the setup-like actions provided by GitHub come with built-in caching functionality, like actions/setup-node, actions/setup-java and others.

Furthermore, there are many examples of community-driven Actions with built-in caching functionality, like ruby/setup-ruby, astral-sh/setup-uv, Swatinem/rust-cache. In general, most of them build on top of actions/toolkit for the sake of easily integrate with GitHub cache server at Workflow runtime.

This vulnerability happens when release workflows leverage build state cached from previous workflow executions, in general on top of the aforementioned actions or similar ones. The publication of artifacts usually happens driven by trigger events like release or events with path filters like push (e.g. for tags).

In such scenarios, an attacker with access to a valid GITHUB_TOKEN can use it to poison the repository's GitHub Actions caches. That compounds with the default behavior of actions/toolkit during cache restorations, allowing an attacker to retrieve payloads from poisoned cache entries, hence achieving code execution at Workflow runtime, potentially compromising ready-to-publish artifacts.

Other resources:

Remediation🔗

In general, you should avoid using previously cached CI state within workflows intended to publish build artifacts:

  • Remove cache-aware actions like actions/cache from workflows that produce releases, or
  • Disable cache-aware actions with an if: condition based on the trigger at the step level, or
  • Set an action-specific input to disable cache restoration when appropriate, such as lookup-only in Swatinem/rust-cache.

secrets-inherit🔗

Type Examples Introduced in Works offline Enabled by default
Workflow secrets-inherit.yml v1.1.0 ✅ ✅

Detects excessive secret inheritance between calling workflows and reusable (called) workflows.

Reusable workflows can be given secrets by their calling workflow either explicitly, or in a blanket fashion with secrets: inherit. The latter should almost never be used, as it makes it violates the Principle of Least Authority and makes it impossible to determine which exact secrets a reusable workflow was executed with.

Remediation🔗

In general, secrets: inherit should be replaced with a secrets: block that explicitly forwards each secret actually needed by the reusable workflow.

reusable.yml
jobs:
  pass-secrets-to-workflow:
    uses: ./.github/workflows/called-workflow.yml
    secrets: inherit
reusable.yml
jobs:
  pass-secrets-to-workflow:
    uses: ./.github/workflows/called-workflow.yml
    secrets:
      forward-me: ${{ secrets.forward-me }}
      me-too: ${{ secrets.me-too }}

bot-conditions🔗

Type Examples Introduced in Works offline Enabled by default
Workflow bot-conditions.yml v1.2.0 ✅ ✅

Detects potentially spoofable bot conditions.

Many workflows allow trustworthy bots (such as Dependabot) to bypass checks or otherwise perform privileged actions. This is often done with a github.actor check, e.g.:

if: github.actor == 'dependabot[bot]'

However, this condition is spoofable: github.actor refers to the last actor to perform an "action" on the triggering context, and not necessarily the actor actually causing the trigger. An attacker can take advantage of this discrepancy to create a PR where the HEAD commit has github.actor == 'dependabot[bot]' but the rest of the branch history contains attacker-controlled code, bypassing the actor check.

Other resources:

Remediation🔗

In general, checking a trigger's authenticity via github.actor is insufficient. Instead, most users should use github.event.pull_request.user.login or similar, since that context refers to the actor that created the Pull Request rather than the last one to modify it.

More generally, GitHub's documentation recommends not using pull_request_target for auto-merge workflows.

bot-conditions.yml
on: pull_request_target

jobs:
  automerge:
    runs-on: ubuntu-latest
    if: github.actor == 'dependabot[bot]' && github.repository == github.event.pull_request.head.repo.full_name
    steps:
      - run: gh pr merge --auto --merge "$PR_URL"
        env:
          PR_URL: ${{ github.event.pull_request.html_url }}
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
bot-conditions.yml
on: pull_request

jobs:
  automerge:
    runs-on: ubuntu-latest
    if: github.event.pull_request.user.login == 'dependabot[bot]' && github.repository == github.event.pull_request.head.repo.full_name
    steps:
      - run: gh pr merge --auto --merge "$PR_URL"
        env:
          PR_URL: ${{ github.event.pull_request.html_url }}
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

overprovisioned-secrets🔗

Type Examples Introduced in Works offline Enabled by default
Workflow overprovisioned-secrets.yml v1.3.0 ✅ ✅

Detects excessive sharing of the secrets context.

Typically, users access the secrets context via its individual members:

env:
  SECRET_ONE: ${{ secrets.SECRET_ONE }}
  SECRET_TWO: ${{ secrets.SECRET_TWO }}

This allows the Actions runner to only expose the secrets actually used by the workflow to the job environment.

However, if the user instead accesses the entire secrets context:

env:
  SECRETS: ${{ toJson(secrets) }}

...then the entire secrets context is exposed to the runner, even if only a single secret is actually needed.

Remediation🔗

In general, users should avoid loading the entire secrets context. Secrets should be accessed individually by name.

overprovisioned.yml
jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
      - run: ./deploy.sh
        env:
          SECRETS: ${{ toJSON(secrets) }}
overprovisioned.yml
jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
      - run: ./deploy.sh
        env:
          SECRET_ONE: ${{ secrets.SECRET_ONE }}
          SECRET_TWO: ${{ secrets.SECRET_TWO }}