langchain-eval-harness
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill establishes an evaluation pipeline that ingests production-sampled data (golden sets) and processes it using LLM-as-judge evaluators (e.g., deepeval GEval). This creates a surface for indirect prompt injection where malicious or adversarial inputs within the dataset could attempt to influence the judge LLM's scoring or behavior.
- Ingestion points: Golden set files (e.g.,
evals/golden_set/v2026.04.jsonl) are read and their contents are passed as inputs to both the target chain and the judge LLM inSKILL.md(Step 2 and Step 4). - Boundary markers: The provided code snippets for custom evaluators and deepeval metrics do not show explicit use of delimiters or instructions to the judge LLM to ignore potential injections within the data being evaluated.
- Capability inventory: The skill utilizes network operations for LLM API calls (OpenAI, Anthropic, LangSmith) and performs file system operations to read datasets and write evaluation results.
- Sanitization: While the instructions recommend redacting PII from production samples, there is no specific guidance on sanitizing inputs to prevent prompt injection attacks targeting the evaluation judge.
Audit Metadata