braintrust-probe-capability-and-variability
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFENO_CODE
Full Analysis
- [SAFE]: The skill consists entirely of instructional Markdown and YAML frontmatter. No executable code, such as scripts or binaries, is present within the skill files.\n- [PROMPT_INJECTION]: Static analysis flagged the string "ignore previous instructions" in
references/interaction-contract.md. Technical review confirms this is a defensive instruction advising the agent to treat data as evidence rather than commands, which is a security best practice for model evaluation.\n- [PROMPT_INJECTION]: The skill outlines procedures for analyzing untrusted trace data. It explicitly addresses the risk of indirect prompt injection inreferences/interaction-contract.mdby requiring the agent to treat such content as evidence and avoid adopting the framing or instructions found within the data.
Audit Metadata