braintrust-attribute-multi-variable-change

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The documentation in references/interaction-contract.md includes a specific security section (Section 9) that identifies common prompt injection patterns like "ignore previous instructions." It instructs the agent to treat trace content and logs as data rather than instructions to prevent directive-following or judgment contamination. This is a documented safety mitigation rather than an attack.
  • [SAFE]: No malicious code, network exfiltration, or unauthorized file access patterns were detected. The skill is entirely composed of markdown files used for procedural guidance in evaluation workflows.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 10:16 PM