flow-specialize
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an inherent vulnerability surface by processing untrusted repository data to generate instructions for other agents. 1. Ingestion points: Processes target tasks, repository instructions, configuration, implementation files, and tests (SKILL.md). 2. Boundary markers: Instructions explicitly require separating rules from precedent, noting local exceptions, and forbid preloading verdicts or ignoring defects. 3. Capability inventory: File system read access for evidence gathering and write access for creating briefs (SKILL.md). 4. Sanitization: No explicit sanitization or escaping of ingested content is defined before interpolation into Markdown briefs.
- [PROMPT_INJECTION]: No patterns of instruction override or safety bypass were found in the skill's own logic.
- [DATA_EXPOSURE_AND_EXFILTRATION]: No hardcoded credentials or network exfiltration patterns were detected.
- [PRIVILEGE_ESCALATION]: The skill explicitly instructs to preserve existing permissions and avoid broadening tool access.
Audit Metadata