Security context

What an agent needs to avoid regressing past fixes and find the next vuln in this repo.

wshobson/agents

              main @ cc37bfd
            

Fixes

CVEs

CRITICAL

Peak severity

Highlights

Auth Bypass: 2 prior fixes. Scrutinize any change in this area.

.github/workflows/claude.yml: most-fixed (1 issue). Treat as high-risk during review.

3 high-severity fixes in this history; regressions here are high-impact.

Recurring patterns

The bug types that recur here, drawn from past fixes, not open vulnerabilities.

Auth Bypass: A logic flaw in the permission block builder allowed an explicit empty tools specification to bypass safety constraints completely, granting full system access instead of restricting it. Developers must explicitly check for empty-list definitions using structural controls.

Auth Bypass: The Claude Code execution workflow did not gate execution permissions on pull requests, allowing unauthorized external actors to run highly privileged commands via actions. Strict author association checks and tool lists are required.

Insecure Workflow Execution: Using standard pull_request triggers on workflows that perform checkout or run scripts on external contributor code opens a direct route to code injection and secret exfiltration. Migrating to pull_request_target with isolated actions is essential.