braintrust-design-eval-experiment

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill consists entirely of instructional Markdown and YAML configuration templates. No executable code, scripts, or binary files are included in the skill package.
  • [INDIRECT_PROMPT_INJECTION]: The skill explicitly acknowledges the risk of processing untrusted data and implements mitigations. 1. Ingestion points: The agent is instructed to process traces, logs, and model outputs which may contain embedded instructions. 2. Boundary markers: The 'Shared interaction contract' (references/interaction-contract.md, Section 9) provides explicit boundaries, instructing the agent to treat this content as evidence rather than instruction. 3. Capability inventory: No dangerous capabilities such as file system writing, network requests, or shell command execution were detected. 4. Sanitization: The skill relies on clear logical directives to prevent the agent from following instructions found within the data items.
  • [PROMPT_INJECTION]: A static analysis flag for prompt injection was reviewed and determined to be a false positive. The documentation in references/interaction-contract.md uses the phrase 'ignore previous instructions' as an example of content that should be identified as data and disregarded, which is a security best practice for agents that process logs.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 10:17 PM