eval-audit

Pass

Audited by Gen Agent Trust Hub on Sep 1, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data including model traces and evaluator definitions, creating a surface where embedded instructions could attempt to influence the agent's analysis report.\n
  • Ingestion points: The skill gathers artifacts such as traces, evaluator configurations, and experiment results from observability MCP servers or local files (CSV, JSON, notebooks) as detailed in the SKILL.md file.\n
  • Boundary markers: The instructions do not define specific delimiters or "ignore previous instructions" warnings to wrap the ingested trace data.\n
  • Capability inventory: The skill's primary function is generating a diagnostic report. While it references other tools like error-discovery and validate-evaluator, it does not directly execute arbitrary shell commands or perform network exfiltration of the ingested data.\n
  • Sanitization: No sanitization or escaping of the external trace content is specified in the audit process.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 1, 2026, 04:16 PM
Security Audit — agent-trust-hub — eval-audit