braintrust-plan-agent-eval
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill includes robust defensive measures against prompt injection. In
references/interaction-contract.md(Section 9), it explicitly instructs the agent to treat trace content, logs, and model outputs as evidence rather than instructions. It specifically warns the agent to ignore strings like "ignore previous instructions" that might appear in the data being evaluated, which is a security best practice for agents processing untrusted corpora. - [DATA_EXFILTRATION]: No patterns of data exfiltration were detected. The skill does not contain instructions to access sensitive files (such as
.envor SSH keys) or perform unauthorized network requests to external domains. - [COMMAND_EXECUTION]: There are no executable scripts or shell command patterns included in the skill. The instructions are purely procedural for planning and metadata inventory within the Braintrust platform environment.
- [SAFE]: The skill demonstrates safe practices by implementing strict boundary markers for data processing and providing clear guidance on how to avoid "judgment contamination" when reading system traces.
Audit Metadata