concept-refine
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and refactor software concepts, rules, caller code, and tests. This introduces a risk where malicious instructions embedded in the source material (such as comments or documentation) could influence the agent's behavior during the 'authorized repair' phase.
- Ingestion points: The workflow in
SKILL.mdexplicitly directs the agent to 'inspect affected concepts, rules, callers, and tests', which are untrusted data sources provided by the user or project environment. - Boundary markers: There are no explicit instructions or delimiters provided to the agent to help it distinguish between structural code to be analyzed and potential natural language instructions embedded within that code.
- Capability inventory: The skill is authorized to 'Complete authorized edits', meaning it has the capability to write or modify files based on its analysis of potentially poisoned input.
- Sanitization: The instructions do not specify any validation or sanitization steps for the ingested code or metadata before the agent acts upon the proposed design changes.
Audit Metadata