debugging-mcp-analytics
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes a pipeline that processes untrusted agent data (intents) which is subsequently summarized by an LLM.
- Ingestion points: Data enters the system via the ClickHouse
eventstable, specifically the$mcp_intentproperty which contains descriptions of agent actions provided by external clients. - Boundary markers: The skill instructions specify that the LLM summary is limited to two sentences and uses structured output for themes, which helps constrain the output, though explicit prompt delimiters for the raw input are not detailed.
- Capability inventory: The process generates LLM-powered summaries and structured theme digests based on these external intent strings, which are then stored in Postgres.
- Sanitization: The documentation in
references/event-vocabulary.mdmentions that captured parameters and responses are sanitized, providing a layer of protection for the data processed by the agent.
Audit Metadata