braintrust-write-agent-behavior-spec

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill includes defensive instructions in references/interaction-contract.md (Section 9) that command the agent to ignore directive-following text (such as 'ignore previous instructions') found within trace data. This is an intentional safety control designed to prevent adversarial data from overriding the agent's primary mission.
  • [INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted data including traces, logs, and datasets, which represents an indirect prompt injection attack surface. However, the risk is mitigated by the lack of powerful executable tools and the presence of explicit instructional boundaries.
  • Ingestion points: Artifacts, traces, and datasets supplied by the user (as referenced in SKILL.md and references/interaction-contract.md).
  • Boundary markers: Explicit instructions in references/interaction-contract.md directing the agent to treat trace content as evidence rather than directives.
  • Capability inventory: The skill is purely instructional and does not possess network access, subprocess execution, or sensitive file-write capabilities.
  • Sanitization: The skill relies on the agent's adherence to metadata-level instructions to differentiate between data observations and actionable instructions.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 10:17 PM