experiment-loop
Pass
Audited by Gen Agent Trust Hub on Mar 31, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill directs the agent to identify or create a 'measurement command' (e.g., a shell script) and execute it autonomously in a loop. This involves running arbitrary local code within the agent's environment to measure performance metrics.
- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through the data it processes during the optimization loop. Malicious instructions embedded in the code being optimized or in the output of benchmark scripts could influence the agent's actions.
- Ingestion points: Implementation files identified for optimization, benchmark scripts within the repository, and the captured output of the measurement command (e.g.,
run.log). - Boundary markers: Absent. The instructions do not provide delimiters or instructions to ignore embedded commands when reading files or parsing benchmark output.
- Capability inventory: The agent is granted
Bash,Edit,Write, andReadtools, providing high-privilege access to the local file system and execution environment. - Sanitization: Absent. The skill does not include steps to sanitize or validate external code or command output before the agent processes it.
Audit Metadata