hyper-decision
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data within its decision-making pipeline. It implements a strong security posture by instructing the agent to treat all input strictly as data and providing adversarial test cases to validate this boundary. Ingestion points include the 'state' field in assets/templates/request.json and the 'input' field in assets/templates/request-openai.json. Boundary markers are established in rules/safety-and-boundaries.md (Section 6), which mandates that the model must ignore any instructions embedded within user input. Capability inventory includes file-writing for generating artifacts and subprocess calls for validation scripts. Sanitization is performed by the scripts/check-decision-request.sh script, which automatically scans request data for potential API keys or secrets.
- [COMMAND_EXECUTION]: The skill provides a bash script (scripts/check-decision-request.sh) and instructs the agent to use it for offline validation of JSON request structures. The script uses standard tools like bash and jq, follows safe scripting practices such as using mktemp for temporary files, and implements secret scanning to prevent accidental credential leakage in request bodies.
- [EXTERNAL_DOWNLOADS]: Documentation and template files reference several well-known technology services and API endpoints, including OpenAI, Vercel, Cloudflare, Perplexity, and TypeSafe. these references are essential to the skill's primary purpose and are documented with clear security boundaries and consent requirements.
Audit Metadata