braintrust-define-eval-release-gate
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The static detector flagged the phrase "ignore previous instructions" within
references/interaction-contract.md. However, this occurs in Section 9, which provides security guidance instructing the agent to treat trace content as data rather than instructions. This is a defensive measure to prevent the agent from following directives embedded in external inputs.\n- [INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection by processing external data such as traces, logs, and datasets.\n - Ingestion points: Data ingested from traces, logs, and dataset rows as described in
references/interaction-contract.mdandreferences/platform-mechanics.md.\n - Boundary markers: Explicitly provided in
references/interaction-contract.md(Section 9), which distinguishes between evidence and instruction.\n - Capability inventory: No scripts or tool definitions within the skill files perform high-risk operations like network calls or arbitrary command execution.\n
- Sanitization: The skill instructions mandate that external content should be quoted as evidence and must not supply its own measurement criteria.
Audit Metadata