What happened
A PR contributor could place a malicious workspace configuration in a repository. When Gemini CLI ran in headless mode with --yolo in CI, it auto-trusted the workspace and ignored tool allowlists, executing attacker-controlled commands on the CI runner before sandbox initialization — with access to all repository secrets.
Why it matters
Arbitrary command execution on CI runners with access to repository secrets (API keys, deployment tokens, signing credentials) and full source code. Any organization using Gemini CLI in CI with headless mode and --yolo was exposed to full CI compromise by any PR contributor.
Missing authorization check
Headless mode must not bypass workspace trust validation. The --yolo flag must not ignore tool allowlists in CI contexts. CI runners require all authorization checks to be enforced before any code from untrusted sources executes.
Would PP block it?
External enforcement means that even if Gemini CLI's internal authorization is bypassed (headless auto-trust, --yolo allowlist skip), the PP enforcement layer still requires a signed receipt before consequential CI actions execute. The bypass of internal controls does not bypass the external gate.