braintrust-deploy-evaluator

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The static analysis detection for instruction override is a false positive. The content in references/interaction-contract.md is a security policy that instructs the agent to ignore malicious directives found in data traces, acting as a defense rather than an attack.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data from logs and traces which presents an indirect injection surface. This is mitigated by a mandatory interaction contract that defines clear boundaries between data and instructions. * Ingestion points: Production logs, traces, and model outputs accessed via sql_query. * Boundary markers: Explicit instructions in references/interaction-contract.md to treat data as evidence only and not as instruction. * Capability inventory: Platform tools for querying, testing, creating, and updating scoring logic (sql_query, test_evaluator, create_evaluator, update_online_scoring_rule). * Sanitization: Strict policy against adopting framing or constraints from evaluated traces.
  • [DYNAMIC_EXECUTION]: The skill facilitates the development of Python and TypeScript scoring functions. This is the intended purpose of the tool and is performed within the secure context of the platform's evaluation environment using provided templates and verification steps.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 10:17 PM