eval-session-classify
Pass
Audited by Gen Agent Trust Hub on Apr 16, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill is a diagnostic tool intended for use within the Datadog ecosystem by the 'datadog-labs' organization. It correctly identifies and uses official Datadog MCP server tools for observability.
- [DATA_EXFILTRATION]: The skill retrieves sensitive data, including user emails and full chat transcripts. However, the use of this data is strictly scoped to the internal classification task within the Datadog platform environment. No external network operations or unauthorized data transfer patterns were detected.
- [INDIRECT_PROMPT_INJECTION]: The skill involves processing untrusted data from session logs and user-uploaded evaluations.
- Ingestion points: User message content and ROUTE_CONTEXT retrieved via 'get_llmobs_agent_loop' (SKILL.md), and reasoning strings from 'get_llmobs_span_details' (SKILL.md).
- Boundary markers: No specific delimiters or warnings to ignore instructions within the retrieved data are implemented.
- Capability inventory: The skill is limited to read operations via MCP tools; it has no capabilities for shell execution, filesystem modification, or arbitrary internet access.
- Sanitization: The instructions do not specify any sanitization or validation for retrieved conversational content.
Audit Metadata