langchain-observability
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill facilitates the setup of standard LangChain observability tools. The provided Python code for MetricCallback and various metric sinks (Prometheus, StatsD, Datadog) follows best practices and contains no malicious logic or obfuscation.
- [SAFE]: Environment variable configuration guidelines emphasize the use of secure secret management. Example API key strings are used solely as format indicators (e.g., lsv2_pt_...) and do not contain real credentials.
- [INDIRECT_PROMPT_INJECTION]: The skill contains a surface for indirect prompt injection as it processes LLMResult objects within its callback handlers.
- Ingestion points: LLM output and metadata are ingested in SKILL.md and references/custom-metrics-callback.md via the on_llm_end callback method.
- Boundary markers: Data is processed as structured objects; no specific text-based delimiters are used for the code-level metadata extraction.
- Capability inventory: The skill performs network operations via the LangSmith client and writes to external metric sinks (Prometheus, StatsD, Datadog).
- Sanitization: The implementation specifically targets structured metadata fields (e.g., input_tokens, output_tokens) rather than raw message content, effectively sanitizing the input against direct injection into the metrics pipeline.
Audit Metadata