exploring-mcp-tool-usage
Pass
Audited by Gen Agent Trust Hub on Sep 26, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data originating from external agent tool calls, which presents an attack surface for instructions embedded in data to influence the agent.
- Ingestion points: Untrusted data enters the agent context via
posthog:query-mcp-tool-sample-intents(retrieving agent intents) andposthog:query-mcp-tool-failures(retrieving error messages from exceptions). - Boundary markers: The instructions do not define explicit delimiters or warnings to ignore instructions embedded within the retrieved intent or error strings.
- Capability inventory: The skill enables high-privilege operations including arbitrary SQL execution via
posthog:execute-sqland detailed tool/session analysis via variousposthog:query-*tools. - Sanitization: There are no specific instructions for the agent to sanitize, validate, or escape the retrieved external content before using it in its reasoning or further tool calls.
Audit Metadata