wjs-looping-feedback

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). This code intentionally wires a GitHub Actions workflow that sends repository contents and issue text to a third‑party LLM action, grants that run the ability to read/edit files and run Bash, and asks the repo owner to store sensitive tokens — creating clear avenues for data exfiltration, remote code execution, and credential misuse if a malicious allowlisted user or the external action is abused.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.95). The workflow ingests outsider-authored free text from a GitHub Issue’s title and body (github.event.issue.title / github.event.issue.body) into the Claude Code prompt at runtime (assets/feedback.yml → “Run Claude Code” → prompt: ... Body: ${{ github.event.issue.body }}), which is then fed into the agent LLM context.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The workflow runs the external GitHub Action anthropics/claude-code-action@v1 (referenced in the "Run Claude Code" step), which is fetched and executed at runtime and performs remote LLM-driven edits based on the supplied prompt, so it is an external runtime dependency that executes code and controls edits.