Actions: Fix actions cache poisoning queries to account for the latest changes in cache write access

[JarLob]

https://github.blog/changelog/2026-06-26-read-only-actions-cache-for-untrusted-triggers/

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates the GitHub Actions cache-poisoning CodeQL queries to align results with GitHub’s default-branch cache write-access rules, reducing findings for low-trust triggers that are now read-only.

Changes:

• Replace “runs on default branch” conditions with hasDefaultBranchCacheWriteAccess(...) across cache-poisoning queries.
• Update/add test workflows and expected outputs to reflect new trigger/write-access behavior.
• Refine dataflow source detection and update query documentation + change note.

Show a summary per file

File	Description
actions/ql/test/query-tests/Security/CWE-349/CachePoisoningViaPoisonableStep.expected	Updates expected results/edges to match new cache write-access filtering.
actions/ql/test/query-tests/Security/CWE-349/CachePoisoningViaDirectCache.expected	Updates expected results/edges to match new cache write-access filtering.
actions/ql/test/query-tests/Security/CWE-349/CachePoisoningViaCodeInjection.expected	Updates expected results to match event-aligned source/sink behavior.
actions/ql/test/query-tests/Security/CWE-349/.github/workflows/cache_write_capable_workflow_dispatch.yml	Adds a workflow_dispatch scenario intended to be cache-write capable in default-branch scope.
actions/ql/test/query-tests/Security/CWE-349/.github/workflows/cache_write_capable_push.yml	Adds a push-to-main scenario intended to exercise code-injection → cache-poisoning logic.
actions/ql/src/change-notes/2026-07-18-read-only-default-branch-cache.md	Documents the analysis behavior change for the three cache-poisoning queries.
actions/ql/src/Security/CWE-349/CachePoisoningViaPoisonableStep.ql	Switches to hasDefaultBranchCacheWriteAccess gating.
actions/ql/src/Security/CWE-349/CachePoisoningViaPoisonableStep.md	Updates explanation/examples to reflect read-only cache access on low-trust triggers.
actions/ql/src/Security/CWE-349/CachePoisoningViaDirectCache.ql	Switches to hasDefaultBranchCacheWriteAccess gating.
actions/ql/src/Security/CWE-349/CachePoisoningViaDirectCache.md	Updates explanation/examples to reflect new write-access rules and provides new unsafe example.
actions/ql/src/Security/CWE-349/CachePoisoningViaCodeInjection.ql	Adds an event-name alignment constraint between flow source and relevant cache-poisoning event.
actions/ql/src/Security/CWE-349/CachePoisoningViaCodeInjection.md	Updates explanation/examples to reflect read-only cache access on low-trust triggers.
actions/ql/lib/codeql/actions/security/CodeInjectionQuery.qll	Updates “relevant cache poisoning event” logic to use cache write-access predicate.
actions/ql/lib/codeql/actions/security/CachePoisoningQuery.qll	Introduces default-branch cache write-event modeling + hasDefaultBranchCacheWriteAccess.
actions/ql/lib/codeql/actions/dataflow/FlowSources.qll	Refactors event-context/json source detection into helper predicates.

Review details

• Files reviewed: 15/15 changed files
• Comments generated: 4
• Review effort level: Low

[Copilot]

Review details

Comments suppressed due to low confidence (6)

actions/ql/src/Security/CWE-349/CachePoisoningViaPoisonableStep.md:65

• The YAML example runs git fetch origin ... / git checkout ... without first checking out or otherwise initializing the repository (no actions/checkout and no git init + git remote add origin ...). As written, the example won’t work on a fresh runner. Add an initial repository checkout/clone step (or explicit git initialization) before the fetch/checkout so the example is runnable and matches the intended scenario.

on:
workflow_dispatch:
inputs:
head_sha:
required: true
jobs:
test:
permissions: {}
runs-on: ubuntu-latest
steps:
- env:
HEAD_SHA: ${{ github.event.inputs.head_sha }}
run: |
git fetch origin "$HEAD_SHA"
git checkout "$HEAD_SHA"
- name: Run tests
run: npm install

actions/ql/src/Security/CWE-349/CachePoisoningViaDirectCache.md:69

• Same issue as the other doc example: this snippet fetches/checks out a SHA without establishing a local git repository or an origin remote first. Add an explicit checkout/clone (or git init + remote setup) step before git fetch to make the example executable and unambiguous.

on:
workflow_dispatch:
inputs:
head_sha:
required: true
jobs:
cache:
permissions: read-all
runs-on: ubuntu-latest
steps:
- env:
HEAD_SHA: ${{ github.event.inputs.head_sha }}
run: |
git fetch origin "$HEAD_SHA"
git checkout "$HEAD_SHA"
- name: Cache fetched files
uses: actions/cache@v4
with:
path: .
key: dispatched-${{ github.event.inputs.head_sha }}

actions/ql/test/query-tests/Security/CWE-349/.github/workflows/cache_write_capable_workflow_dispatch.yml:14

• Setting permissions: {} at the job level removes all GITHUB_TOKEN permissions, which can break actions/checkout (typically requires at least contents: read). Even though this is a test fixture, it’s helpful for fixtures to be runnable/representative—consider specifying minimal required permissions (for example contents: read) instead of an empty map.

 cache:
permissions: {}
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4

actions/ql/lib/codeql/actions/security/CachePoisoningQuery.qll:51

• The helper returns multiple event names (a set via multiple result bindings), but the name defaultBranchCacheWriteEvent reads like a single value. Renaming to something plural (for example defaultBranchCacheWriteEvents) would make the intent clearer and reduce confusion with defaultBranchTriggerEvent().

private string defaultBranchCacheWriteEvent() {
result =
[
"push", "workflow_dispatch", "repository_dispatch", "delete", "registry_package",
"page_build", "schedule"
]
}

actions/ql/src/Security/CWE-349/CachePoisoningViaCodeInjection.ql:28

• This uses a type-cast source.getNode().(RemoteFlowSource) as a filter. If the CodeInjection flow config ever admits non-RemoteFlowSource nodes as source, this will silently drop those paths. Consider rewriting this constraint using an explicit exists(RemoteFlowSource rfs | rfs = source.getNode() and ...) (or equivalent) to make the filtering intent explicit and more robust to future changes.

where
CodeInjectionFlow::flowPath(source, sink) and
event = getRelevantCachePoisoningEventForSink(sink.getNode()) and
source.getNode().(RemoteFlowSource).getEventName() = event.getName() and
// the checkout is not controlled by an access check
not exists(ControlCheck check |
check.protects(source.getNode().asExpr(), event, "code-injection")

actions/ql/src/Security/CWE-349/CachePoisoningViaPoisonableStep.ql:43

• hasDefaultBranchCacheWriteAccess(job, event) already includes job.getATriggerEvent() = event. Keeping the separate job.getATriggerEvent() = event constraint here is redundant and makes the condition harder to read/maintain. Consider removing the duplicate constraint (or changing hasDefaultBranchCacheWriteAccess to only take job and derive the event internally).

 job.getATriggerEvent() = event and
// job can be triggered by an external user
event.isExternallyTriggerable() and
hasDefaultBranchCacheWriteAccess(job, event) and

• Files reviewed: 15/15 changed files
• Comments generated: 0 new
• Review effort level: Low

[Copilot]

Review details

Comments suppressed due to low confidence (4)

actions/ql/src/Security/CWE-349/CachePoisoningViaCodeInjection.ql:25

• This hard cast to RemoteFlowSource will silently drop any paths where the source node is not a RemoteFlowSource, potentially introducing false negatives. Consider rewriting this as an exists(RemoteFlowSource rfs | rfs = source.getNode() and rfs.getEventName() = event.getName()) constraint (or guarding with an instanceof check) so the query remains robust to additional source kinds.

 source.getNode().(RemoteFlowSource).getEventName() = event.getName() and

actions/ql/test/query-tests/Security/CWE-349/.github/workflows/cache_write_capable_workflow_dispatch.yml:14

• permissions: {} removes contents: read, which actions/checkout requires to fetch repository contents; as written, this fixture workflow would fail if executed. Even for a test fixture, it’s helpful to keep examples runnable—consider setting minimal permissions (for example contents: read) or removing the explicit empty permissions block.

 cache:
permissions: {}
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4

actions/ql/src/Security/CWE-349/CachePoisoningViaPoisonableStep.md:63

• The ‘Incorrect Usage’ example runs git fetch/checkout without first checking out (or initializing) a repository, so it won’t work as a standalone workflow snippet for readers. To keep the example runnable while still demonstrating the vulnerability, add an initial actions/checkout step (or git init + git remote add origin ...) before the fetch.

 steps:
- env:
HEAD_SHA: ${{ github.event.inputs.head_sha }}
run: |
git fetch origin "$HEAD_SHA"
git checkout "$HEAD_SHA"

actions/ql/src/Security/CWE-349/CachePoisoningViaDirectCache.md:64

• Similarly here, the example uses git fetch/checkout without first checking out or initializing a repository, which makes the snippet non-functional for readers. Add an actions/checkout step (or equivalent repo initialization) before the git commands so the example can be followed as-is.

 steps:
- env:
HEAD_SHA: ${{ github.event.inputs.head_sha }}
run: |
git fetch origin "$HEAD_SHA"
git checkout "$HEAD_SHA"

• Files reviewed: 15/15 changed files
• Comments generated: 0 new
• Review effort level: Low

[redsun82]

Thanks!! Overall looks good to me, with just one testing gap that could be filled

[JarLob]

Implemented suggestions and added tests. Also noticed one incorrect assumption about workflow_call

[redsun82]

LGTM, thanks!

[JarLob]

Thanks! Please merge, I don't have permissions.