Supply Chain Security

GitHub Cache Modes Can Restrict Access or Override a Safer Default

By Kaleido Field Staff · September 11, 2026

Citation-ready: GitHub Actions cache-mode can override low-trust events' read-only default; write access triggers a warning and can increase cache-poisoning risk.

Evidence boundary: Official configuration documentation. No exploit test or security assessment of a user repository; the article does not provide an attack procedure.

What happened and why it matters

The new control is useful only when its overrides are reviewed against the event's trust level; a configuration change is not automatically a security improvement.

Primary evidence

Primary reference: GitHub Changelog. Kaleido Field checked the event date and the article's attributed facts against this source.

Read and write are distinct cache operations

The modes are read, write, write-only and none. Write permits both restores and saves; write-only permits saves without restores. Job settings override workflow settings.

Choose the smallest scope that the job needs. A publishing job that never consumes a cache has different requirements from a test job that only restores dependencies. Those examples describe access needs, not a universal configuration prescription.

A called workflow cannot gain more than its caller grants

GitHub says enforcement happens at the cache service and carries through reusable workflows. The caller's grant sets a ceiling for the called workflow. Existing defaults remain when cache-mode is not declared.

Review the caller and callee together. A local YAML line can be misleading if the actual run inherits a stricter grant from its caller or a different event context than the reviewer assumed.

Treat the warning as an unresolved decision

For low-trust events such as pull_request_target, an explicit write or write-only mode overrides the read-only default and generates a warning. Do not treat the warning annotation as proof that unsafe behavior was prevented.

Keep that review alongside the AI Scan configuration record. Both controls benefit from an effective-state readback and a separate observation of the run they are meant to govern.

Evidence boundary

Official configuration documentation. No exploit test or security assessment of a user repository; the article does not provide an attack procedure.