agent-evals
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFENO_CODE
Full Analysis
- [SAFE]: The skill is composed entirely of documentation (Markdown files) and architectural references. It does not include any executable scripts, automated installation routines, or shell commands that could perform malicious actions.\n- [SAFE]: External links reference highly reputable sources in the AI safety and research community, such as Anthropic and METR, for legitimate educational and benchmarking purposes.\n- [SAFE]: The documentation demonstrates a high degree of security awareness by warning users about potential data exposure in third-party telemetry tools (e.g., OpenLLMetry) and recommending privacy redaction for production logs.\n- [INDIRECT_PROMPT_INJECTION]: The skill describes an evaluation framework that ingests untrusted data from production traces and user transcripts. While this creates an attack surface for indirect prompt injection against the evaluator, the skill mitigates this by recommending human-in-the-loop calibration and narrow, deterministic checks. Ingestion points: Production traces and user transcripts (SKILL.md Section 6). Boundary markers: Not explicitly implemented in the markdown instructions. Capability inventory: No execution capabilities (subprocess, eval) are present in the skill files. Sanitization: Recommends conversation redaction before processing (support-quality.md A1).
Audit Metadata