skill-trace-evaluation

Pass

Audited by Gen Agent Trust Hub on Jun 23, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes a local Python script (scripts/trace_evaluate.py) to analyze file structures and content patterns in a target skill directory. This execution is confined to the local environment and does not involve elevated permissions or network connectivity.- [PROMPT_INJECTION]: An indirect prompt injection surface is present because the skill is designed to read and interpret the contents of external SKILL.md and reference files. Adversarial content in those files could attempt to influence the agent's evaluation logic. Mitigation measures are provided in the instructions, such as a strict calibration limit of ±0.3 and a requirement for specific evidence for every score adjustment.
  • Ingestion points: SKILL.md and documentation within the references/ directory of the skill under evaluation.
  • Boundary markers: Absent. The AI is directed to read ingested content directly to perform semantic calibration.
  • Capability inventory: Execution of a local Python script and the ability to write evaluation reports to the local file system.
  • Sanitization: No specific sanitization or escaping is performed on the external content before it is processed by the agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 23, 2026, 04:05 PM
Security Audit — agent-trust-hub — skill-trace-evaluation