deep-grill
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill implements a robust interrogation workflow consisting of clarification and adversarial phases to improve project specs. The logic is focused on interactive dialogue and report generation. Evidence: SKILL.md workflow and Phase 1/2 descriptions.\n- [INDIRECT_PROMPT_INJECTION]:\n
- Ingestion points: The skill ingests user-provided plans or designs and content from the local codebase (SKILL.md Step 2).\n
- Boundary markers: No explicit boundary markers or "ignore" instructions are specified for external content.\n
- Capability inventory: The skill can read local files for exploration and write decision record files to the
docs/local/deliverables/deep-grill/directory.\n - Sanitization: No specific sanitization or filtering logic is mentioned for processed data.\n
- Analysis: While the skill processes untrusted input, its capabilities are restricted to documentation generation. The absence of network operations or shell execution limits the potential impact of an injection attack.\n- [DATA_EXPOSURE]: The skill is instructed to explore the local codebase to answer questions automatically rather than asking the user (SKILL.md Step 2). This involves read access to the repository, but the information is used solely for generating the local decision record, and there are no patterns suggesting exfiltration.
Audit Metadata