braintrust-discover-trace-topics
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill includes defensive instructions to protect against prompt injection. In
references/interaction-contract.md, it explicitly warns the agent that trace content is evidence and should not be treated as instruction, specifically citing phrases like "ignore previous instructions" as examples of malicious patterns to be disregarded. The static detection of these phrases is a false positive, as they are used for training defensive behavior. - [DYNAMIC_EXECUTION]: The skill guides the agent in generating JavaScript preprocessors and scorers. This is the primary function of the tool and is executed within the platform's execution environment (QuickJS and Node.js). The code generation follows static templates for evaluation purposes.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of trace traffic.
- Ingestion points: Untrusted data enters the context via
project_logsanddataset rowsdefined inSKILL.mdandreferences/facet-pipeline.md. - Boundary markers: Robust boundary markers are present in
references/interaction-contract.md(Section 9), directing the agent to maintain a strict separation between processed data and its own instructions. - Capability inventory: The skill uses platform-managed
inline_functionexecution; it does not have generic file-write or network capabilities outside the Braintrust toolset. - Sanitization: The skill implements regex-based normalization via
no_match_patternto sanitize model outputs before they reach the scoring stage.
Audit Metadata