taste-driven-designer
Pass
Audited by Gen Agent Trust Hub on Sep 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill includes a local utility script,
scripts/generate-seed.sh, designed to generate random seed strings. The agent is instructed to execute this script locally to inject randomness into the design process. This is a core feature and uses standard system utilities. - [INDIRECT_PROMPT_INJECTION]: The framework ingests seed strings as creative input.
- Ingestion points: Random strings from the
generate-seed.shscript are processed in the 'Discover' phase. - Boundary markers: The instructions include a specific guideline in
templates/seed-string-procedure.mdto prevent the agent from reflecting the seed string in final design outputs. - Capability inventory: The skill uses local shell execution and instructs the agent to use external multimodal APIs if configured.
- Sanitization: Input is limited to the output of the provided random generator script.
- [EXTERNAL_DOWNLOADS]: The documentation discusses the use of well-known external multimodal services (such as OpenAI, Runway, and Luma) for asset generation. These are treated as optional user-configured capabilities rather than automated or unauthorized downloads.
Audit Metadata