hypothesis-diversification
Pass
Audited by Gen Agent Trust Hub on Jul 23, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [SAFE]: The skill consists entirely of natural language instructions and configuration for hypotheses generation. No malicious code, obfuscation, or unauthorized data access patterns were detected across the SKILL.md, SKILL_CARD.md, or agents/openai.yaml files.
- [PROMPT_INJECTION]: The skill processes untrusted input such as research questions and plans, which constitutes an indirect prompt injection surface. However, the risk is mitigated by explicit instructions for an 'adversarial pass', 'validation gates', and manual reviewer handoffs before any proposed edits are applied to artifacts.
- Ingestion points: Inputs include research questions, decisions, plans, and SkillOpt targets as defined in SKILL_CARD.md.
- Boundary markers: Not explicitly defined in the prompt text using technical delimiters, but instructions suggest using adversarial notes and rejected candidates to distinguish candidate sets.
- Capability inventory: The skill instructs the agent to propose 'append', 'replace', 'delete', or 'move' operations for target artifacts in SKILL.md.
- Sanitization: Proposals are subject to structural validation and manual review gates as specified in SKILL_CARD.md.
Audit Metadata