dialectic
Pass
Audited by Gen Agent Trust Hub on Sep 24, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied ideas, text, and code to generate refined expressions and alternatives. This ingestion of untrusted data creates a surface for indirect prompt injection where instructions embedded in the user input could attempt to redirect agent behavior.\n
- Ingestion points: User prompts and feedback loops defined in the interaction methodology (SKILL.md).\n
- Boundary markers: The skill suggests using headings, labels (e.g., "A", "B"), and dedicated code blocks to separate articulations, which helps maintain context (references/conversations.md).\n
- Capability inventory: The skill produces prose, code sketches, and rendered previews; it does not define automated network or file-system write operations.\n
- Sanitization: The instructions do not specify explicit sanitization or validation for user-provided fragments before they are synthesized into new attempts.
Audit Metadata