agent-observability
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a 'Virtuous Cycle' workflow that ingests production anomalies and trace data to improve evaluation datasets, creating an entry point for untrusted content.
- Ingestion Points: Production traces, tool outputs, and reasoning logs (SKILL.md).
- Boundary Markers: The skill explicitly instructs agents to 'require schema validation at every boundary' and use 'opt-in sanitized message attributes'.
- Capability Inventory: The frontmatter allows tools for file writing and command execution, which could be targeted by poisoned trace data if not properly handled.
- Sanitization: Instructions mandate that 'the input/output pair is sanitized and added to the golden dataset' before use in the evaluation loop.
Audit Metadata