jmb-triage

Fail

Audited by Gen Agent Trust Hub on Nov 19, 2026

Risk Level: HIGHCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructions for 'Step 3: Verify the claim' mandate that the agent check out and run 'relevant tests or commands' for external Pull Requests. Since these PRs originate from external, untrusted sources, this behavior allows for arbitrary command execution on the agent's host environment via malicious test scripts or modified build configurations (e.g., poisoned package.json scripts or Makefiles).
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from issue bodies, comments, and PR diffs to generate summaries and 'grill' reporters. Maliciously crafted issue content could attempt to influence the agent's logic or exfiltrate data through the issue tracker.
  • Ingestion points: GitHub/GitLab issue bodies, reporter comments, and Pull Request diffs (referenced in SKILL.md).
  • Boundary markers: The skill requires a disclaimer for its own output but does not define markers or 'ignore' instructions for the ingested untrusted content.
  • Capability inventory: Shell command execution (running tests), file system write access (creating/modifying .out-of-scope/*.md files), and network operations via the issue tracker API.
  • Sanitization: No sanitization, validation, or escaping of the external issue/PR content is specified before it is used to inform agent actions or executed as code.
  • [REMOTE_CODE_EXECUTION]: The requirement to 'verify' external PRs by executing their included code constitutes a remote code execution vector. An attacker can achieve RCE by submitting a PR with a malicious payload in the test suite that the agent is instructed to run automatically.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Nov 19, 2026, 07:02 PM