evolve
Warn
Audited by Gen Agent Trust Hub on Mar 10, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill accepts an arbitrary shell command as the 'fitness_cmd' argument and executes it directly on the host system to evaluate code candidates. Evidence: The instruction documentation and shell snippets show the use of 'output=$($FITNESS_CMD 2>&1)' during both the initialization and evaluation phases.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it incorporates local source file contents into prompts for the AI agents responsible for breeding new code. 1. Ingestion points: Target source files specified via the '--files' argument or user prompt. 2. Boundary markers: The prompt templates use placeholders like '[file contents]' but do not include explicit instructions for the AI to ignore embedded commands or comments. 3. Capability inventory: The skill has the ability to create git branches, write files to the repository, and execute shell commands. 4. Sanitization: No sanitization or filtering is performed on the code files before they are provided to the LLM breeder.
Audit Metadata