skill-evals
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONNO_CODE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes a 'god mode' improvement loop where an agent is tasked with analyzing execution traces of another agent. Because these traces contain logs of all tool calls and messages—which may include data from untrusted external sources—they represent a vector for indirect prompt injection if the analyzing agent processes them as instructions rather than data.
- Ingestion points: Execution traces containing logs of tool calls, messages, and results (SKILL.md).
- Boundary markers: The instructions do not specify the use of delimiters or warnings to ignore instructions embedded within the traces.
- Capability inventory: The analyzing agent identifies failure patterns and suggests fixes or new evaluation tasks (SKILL.md).
- Sanitization: No sanitization or validation of the trace content is mentioned in the proposed workflow.
Audit Metadata