langsmith-graph-evaluator
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [Indirect Prompt Injection Surface]: The skill creates evaluators that ingest data from application runs (inputs and outputs) and interpolate them directly into LLM prompts. This creates a surface where potentially adversarial content within the processed data could influence the evaluator's judgment.
- Ingestion points: Untrusted data enters the agent context via the
Run.inputs,Run.outputs, andthread.runsobjects inSKILL.md. - Boundary markers: The provided prompt templates use direct variable interpolation (e.g.,
{{answer}}) without explicit boundary markers or instructions to ignore embedded commands. - Capability inventory: The skill executes Python logic within
@nodefunctions and performs network operations to interface with LLM providers. - Sanitization: The instructions do not describe explicit sanitization, escaping, or filtering of the external content before it is interpolated into prompts.
- [Dependency Management]: The skill guides the user to install external Python packages like
openevalsandlangchain-anthropic. These are standard dependencies for building evaluators within this ecosystem. - [Credential Safety Practices]: The skill provides instructions for managing API keys using environment variables and workspace secrets. It follows a safe practice by explicitly advising against hardcoding sensitive credentials in source files.
Audit Metadata