sa-improve
Fail
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION] (HIGH): The skill's core function is to analyze untrusted user-supplied code, making it highly vulnerable to indirect prompt injection where malicious instructions are hidden within the code or logs.
- Ingestion points: Processes 'Simulated Annealing (SA) solver/code snippet' and 'score/logs' provided by the user.
- Boundary markers: Absent; there are no instructions to the agent to isolate the untrusted code from its command context.
- Capability inventory: High; the skill references an external shell script (scripts/sa_improve.sh), which indicates high-privilege shell execution capabilities.
- Sanitization: Absent; no filtering or validation of user-provided content is mentioned.
- [COMMAND_EXECUTION] (MEDIUM): The skill relies on a local shell script for its operations. Executing logic via shell scripts based on analysis of untrusted input presents a danger of command injection if the script processes the input unsafely.
- Evidence: Instruction to 'Read scripts/improve.sh' suggests dependency on executable shell logic.
Recommendations
- AI detected serious security threats
Audit Metadata