langchain-agent-experiments
Pass
Audited by Gen Agent Trust Hub on Nov 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- Indirect Prompt Injection Surface: The skill is designed to ingest and process data from external sources, such as JSON datasets and production traces. This creates a potential surface for indirect prompt injection if the data contains malicious instructions. The skill proactively mitigates this by instructing the agent to treat this data as untrusted and explicitly telling LLM judges to treat inputs as data rather than instructions. Evidence of this pattern is found in the instructions for building datasets and writing evaluators in SKILL.md.
- Command Execution via CLI: The skill utilizes the
langsmithcommand-line interface for managing datasets, experiments, and evaluators. These operations are essential for the skill's workflow and are performed using the authenticated user's permissions. Examples includelangsmith datasets uploadandlangsmith experiments list. - Dynamic Code Execution: The workflow involves defining and running evaluator functions in Python or TypeScript to grade application outputs. This is a core feature of the experimentation process but involves running code that may be dynamically generated or modified. The skill notes that server-side evaluators run within a restricted sandbox environment.
- External Resource Integration: The skill references several external packages and CLI tools, specifically
langsmith,openevals, andagentevals. These are official resources provided by the vendor for the intended purpose of evaluation and monitoring and are handled according to standard development practices. - Secure Distributed Tracing: The instructions for tracing deployed agents into experiments (documented in references/remote-agent-tracing.md) involve passing headers like
langsmith-traceandbaggage. The skill correctly identifies the security risks and advises that these headers should only be applied for trusted internal callers and stripped from public traffic.
Audit Metadata