design-critic-loop
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill instructions define a legitimate multi-phase workflow for design quality assurance without the use of dangerous commands or unauthorized access.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data, specifically design artifacts (code, HTML, images) and reference URLs provided by the user. While this creates a surface for indirect prompt injection, the skill mitigates this through its architecture.
- Ingestion points: The
ARTIFACTandREFERENCEvariables processed during Phase 3. - Boundary markers: The subagent prompts in
references/critic-prompts.mduse clear headers (e.g., 'The artifact:', 'The brief:') to separate instructions from untrusted data. - Capability inventory: The process is restricted to generating textual critiques via the
Agenttool; no shell execution, file-writing, or network capabilities are present in the scripts. - Sanitization: The workflow employs 'fresh-context' subagents to isolate the evaluation process, preventing any malicious instructions in the artifact from influencing the main agent's session.
Audit Metadata