discovering-high-level-patterns
Fail
Audited by Gen Agent Trust Hub on Jun 22, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill's workflow (Steps 3 and 5) explicitly instructs the agent to synthesize Python functions (e.g.,
def detect_pattern(...)) and then execute them over raw trace data. Generating and running code dynamically is a high-risk pattern as it can lead to arbitrary code execution. - [COMMAND_EXECUTION]: The skill describes an iterative loop (Step 4) for evaluating and refining synthesized programs. This includes catching runtime errors and feeding them back into the LLM for 'repair,' confirming that a live code execution environment is a requirement for this skill.
- [PROMPT_INJECTION]: The skill possesses a broad surface for indirect prompt injection. Malicious instructions could be embedded within the simulation logs or natural language goals (Category 8), which are then ingested as context for the code synthesis process, potentially causing the agent to generate and execute harmful scripts.
- Ingestion points: Simulation snapshots
tau = [x_1, x_2, ..., x_T]and natural language goal descriptions. - Boundary markers: None present; raw data is processed into state dictionaries.
- Capability inventory: Ability to generate and execute Python code via a synthesis loop.
- Sanitization: No sanitization or sandboxing of the synthesized code is mentioned in the workflow.
Recommendations
- AI detected serious security threats
Audit Metadata