gan-style-harness
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONMETADATA_POISONING
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill architecture creates a feedback loop where an Evaluator agent interacts with a live application created by a Generator agent.\n
- Ingestion points: The Evaluator agent ingests content from a live application via Playwright interaction; the Generator agent ingests feedback files (e.g., feedback-001.md) produced by the Evaluator.\n
- Boundary markers: The instructions suggest passing feedback via files rather than inline, but provide no specific delimiters or instructions to ignore malicious content within the live application being tested.\n
- Capability inventory: The harness uses high-privilege tools including Bash, Write, and Playwright for live system interaction.\n
- Sanitization: No sanitization or validation of the live application's output is described before processing by the Evaluator.\n- [DYNAMIC_EXECUTION]: The skill facilitates the generation and runtime execution of arbitrary application code.\n
- The instructions involve starting a development server (npm run dev) and running generated code to facilitate live testing.\n- [METADATA_POISONING]: The skill contains deceptive references to non-existent documentation and technology.\n
- The description and references cite a fictional Anthropic engineering paper from March 2026 and Opus 4.6 models, which may provide a false sense of authority or technical grounding.
Audit Metadata