deep-think
Pass
Audited by Gen Agent Trust Hub on May 14, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface due to the direct interpolation of untrusted content into its reasoning templates.
- Ingestion points: Variables such as
{problem},{context},{stuck_reason}, and{attempts}are taken from the user and inserted into prompt templates defined inSKILL.md,references/decompose-prompt.md, andreferences/unstick-prompt.md. - Boundary markers: No delimiters or isolation markers (e.g., XML tags or triple backticks) are used to wrap user-provided data, nor are there instructions to the sub-agents to ignore commands embedded in those inputs.
- Capability inventory: The skill has the capability to spawn and coordinate multiple high-reasoning agents (using
model="opus") which could be misdirected by adversarial input. - Sanitization: There is no evidence of input validation, escaping, or sanitization performed on the ingested variables.
Audit Metadata