langchain-eval-engineering
Pass
Audited by Gen Agent Trust Hub on Nov 20, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- Indirect Prompt Injection Surface: The skill is designed to analyze untrusted data, such as agent traces and repository content. This creates an attack surface where instructions hidden in the processed data could attempt to override agent behavior. The skill mitigates this by providing guidance on using boundary markers and instructing evaluation judges to disregard instructions within the analyzed content, following industry best practices for secure data handling.
- Dynamic SQL Execution: The
scripts/snapshot_sqlite_state.pyutility executes SQL queries from a configuration file. Dynamic execution of SQL can pose risks if not properly constrained. The utility implements a restrictive SQLite authorizer to prevent data modification (INSERT, UPDATE, DELETE) and limits queries toSELECTandWITHstatements, ensuring that the operation remains read-only and secure. - External Command Execution: The framework uses command-line tools like
langsmithfor trace discovery andharborfor running evaluation tasks. Executing external commands involves interacting with the underlying system and potentially external networks. The skill manages this risk by providing clear instructions on secure credential management, environment isolation (using Docker), and following the principle of least privilege.
Audit Metadata