braintrust-validate-eval-scorer
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill includes robust defensive instructions in
references/interaction-contract.md(§9) to mitigate potential prompt injection attempts. It explicitly directs the agent to treat content from traces, logs, and datasets as evidence rather than instructions. It specifically warns the agent to ignore "directive-following" attempts (such as "ignore previous instructions") that might be embedded in the data being evaluated. - [INDIRECT_PROMPT_INJECTION]: The skill documentation acknowledges an attack surface where untrusted data is processed, but it establishes clear security boundaries.
- Ingestion points: Untrusted data enters the agent context through traces, model outputs, and dataset rows mentioned in
SKILL.mdandreferences/interaction-contract.md. - Boundary markers: The skill establishes a strict conceptual boundary by mandating that trace content be treated as evidence and never adopted as instruction or criteria.
- Capability inventory: The skill instructions do not include network operations, file system writes, or subprocess execution that could be exploited via injection.
- Sanitization: Logical sanitization is achieved through the requirement to quote trace content as evidence and maintain independent criteria.
Audit Metadata