gan-style-harness
Pass
Audited by Gen Agent Trust Hub on Jun 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The documentation describes workflows that use the
Bashtool to start local development servers (e.g.,npm run dev) and execute shell scripts (e.g.,./scripts/gan-harness.sh). While standard for software development, these operations execute with the agent's privileges. - [PROMPT_INJECTION]: The skill architecture creates an indirect prompt injection surface by establishing a feedback loop where different agents read from and write to shared files such as
spec.mdandfeedback-NNN.md. Malicious instructions embedded in the initial user prompt or generated code could potentially influence the behavior of other agents in the chain. - Ingestion points:
spec.md,feedback-NNN.md(as described in the workflow instructions) - Capability inventory:
Bash,Write,Edit,Read,Task, andPlaywright MCPtools - Boundary markers: None explicitly mentioned to distinguish between agent instructions and untrusted data in the shared files
- Sanitization: The guide does not specify methods for sanitizing or validating the content exchanged between the Planner, Generator, and Evaluator agents
Audit Metadata