braintrust-build-eval-dataset

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and processing untrusted data, such as production traces, logs, and model outputs, which are known vectors for indirect prompt injection. However, the skill provides comprehensive defensive guidance to mitigate these risks.
  • Ingestion points: Production traces and logs are identified as primary data sources in SKILL.md and references/dataset-construction.md.
  • Boundary markers: references/interaction-contract.md (Section 9) provides explicit instructions to treat trace content as evidence rather than instruction, specifically warning the agent to ignore directives like "ignore previous instructions" found within the data.
  • Capability inventory: The skill is scoped to dataset construction, versioning, and metadata management. It does not include scripts for arbitrary command execution or unauthorized network access.
  • Sanitization: The instructions mandate human review queues for production-to-dataset pipelines and PII scrubbing, ensuring a layer of manual verification for external data.
  • [PROMPT_INJECTION]: Static analysis identified a potential prompt injection pattern in references/interaction-contract.md. Upon manual review, this is a false positive. The text in question is a defensive instruction advising the agent to ignore malicious directives found within ingested data traces, which enhances the security posture of the skill.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 10:16 PM