braintrust-design-eval-metric-bundle

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: Static analysis identified potential injection patterns such as 'ignore previous instructions' within references/interaction-contract.md. Analysis confirms this is a false positive; the text is part of a security instruction that warns the agent specifically not to follow such commands when they appear in behavioral traces.
  • [INDIRECT_PROMPT_INJECTION]: The skill acknowledges that analyzing system traces and logs introduces a vulnerability surface where untrusted data enters the agent's context. It proactively mitigates this risk by establishing clear boundary markers in Section 9 of the interaction contract, stating that trace content is evidence and must not be adopted as instructions.
  • [DATA_EXFILTRATION]: No patterns of unauthorized data access or exfiltration were found. References to Braintrust API endpoints (e.g., experiment deletion) are provided as documentation for manual platform hygiene and are not invoked by the skill itself.
  • [NO_CODE]: The skill consists entirely of Markdown and YAML configuration files with no accompanying scripts, binary files, or executable dependencies, significantly reducing the potential attack surface.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 10:16 PM