agent-evolution-patterns
Pass
Audited by Gen Agent Trust Hub on Apr 7, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill facilitates 'Indirect Prompt Injection' by design. It captures instructions (instincts) from session data and injects them into future agent prompts via overlay files.
- Ingestion points: Untrusted session data, such as code under review or external information processed during a session, serves as the source for captured instincts.
- Boundary markers: The skill uses delimiters like '--- Agent Learned Instincts ---' when injecting content into the session context, but these do not prevent the agent from being influenced by adversarial content within those blocks.
- Capability inventory: The system writes persistent markdown files to the user's home directory ('~/.clarc/agent-instincts/') and executes Node.js scripts to manage and load these instructions into every new session.
- Sanitization: The skill relies on a manual approval workflow ('/agent-evolution') where a user must explicitly approve, skip, or edit candidate instincts before they are promoted to the persistent store.
- [COMMAND_EXECUTION]: The skill documentation and functional logic rely on the execution of local scripts ('node scripts/agent-evolution.js') to perform file operations and manage the evolution pipeline.
Audit Metadata