case-outcome-predictor
Warn
Audited by Gen Agent Trust Hub on Mar 23, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The instruction to 'Investigate the entire codebase thoroughly' creates a significant surface for indirect prompt injection. If the codebase being audited contains malicious instructions (e.g., in comments or string literals), the agent could be tricked into deviating from its assigned task.
- [DATA_EXFILTRATION]: The 'SELF-EVOLUTION TELEMETRY' section directs the agent to append metadata to '~/.claude/projects/skill-telemetry.md'. Writing to paths outside the immediate project scope, especially within a hidden application configuration directory, constitutes a persistence and tracking mechanism that stores session data across different environments.
- [PROMPT_INJECTION]: The skill uses a 'SELF-HEALING VALIDATION' loop that can run up to two iterations. This loop-based logic, when processing untrusted input from a codebase, could be manipulated by an attacker to cause excessive computation or specific state changes.
Audit Metadata