grinding-until-pass
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to "capture the output" of external commands (e.g., npm test) and "analyze and fix" based on that output. This creates a surface for indirect prompt injection where malicious content in test logs, error messages, or build outputs could influence the agent's code modifications.
- Ingestion points: Command output captured in Step 2 of SKILL.md.
- Boundary markers: Absent. The instructions do not tell the agent to treat output as untrusted or use delimiters.
- Capability inventory: The agent is granted the ability to execute shell commands (npm, npx, bash) and modify source code files (Step 3: "Make the minimal fix").
- Sanitization: Absent. There are no instructions for filtering or escaping the captured output before processing.
- [COMMAND_EXECUTION]: The skill's primary function is to execute shell commands provided as "goal commands" (e.g., npm test, npm run build). While these are standard development tasks, the autonomous nature of the loop (up to 10 iterations) increases the impact of any underlying vulnerabilities in the scripts being run, especially if they are influenced by the automated fixes.
Audit Metadata