braintrust-elicit-eval-criteria

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data including traces, logs, and datasets (ingestion points in SKILL.md and references/interaction-contract.md). While this presents a surface for injection, references/interaction-contract.md (Section 9) establishes clear boundary markers and instructional sanitization by directing the agent to treat such content as evidence rather than instruction. The skill's capabilities are limited to analysis and reporting, with no dangerous file-write or network operations identified.
  • [PROMPT_INJECTION]: Heuristic detection of instruction overrides was found to be defensive in nature. The instructions in references/interaction-contract.md advise the agent to disregard 'ignore previous instructions' markers found within analyzed traces, rather than attempting to hijack agent behavior.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 10:17 PM