braintrust-design-human-eval-review
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFENO_CODE
Full Analysis
- [PROMPT_INJECTION]: The skill includes phrases such as "ignore previous instructions" within its interaction contract. However, these are strictly documented as examples of malicious patterns the agent should recognize and ignore when reviewing external data traces (Section 9 of interaction-contract.md). This is a proactive security defense mechanism rather than a functional injection.
- [INDIRECT_PROMPT_INJECTION]: The skill defines workflows for processing external data including "traces," "logs," and "production samples."
- Ingestion points: Data traces and dataset rows enter the context for evaluation as described in interaction-contract.md.
- Boundary markers: The skill implements a robust interaction contract that explicitly instructs the agent to treat trace content as evidence only and not as instruction, effectively mitigating injection risks.
- Capability inventory: The skill is purely instructional and does not utilize any shell tools, network requests, or file-writing capabilities.
- Sanitization: Instructions require the agent to quote trace content as evidence and forbid adopting the evaluated system's internal framing or directives.
- [REMOTE_CODE_EXECUTION]: No code or external scripts are included in or downloaded by the skill. All instructions are provided in static Markdown files.
Audit Metadata