openai-responses-adapter
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFENO_CODEINDIRECT_PROMPT_INJECTION
Full Analysis
- [NO_CODE]: The skill package is composed entirely of Markdown documentation, technical specifications, and a license file. It contains no executable scripts, binaries, or configuration files that would be executed by the agent environment, presenting no direct code execution risk.
- [INDIRECT_PROMPT_INJECTION]: The skill describes a protocol for translating untrusted external request data (OpenAI Responses API) for upstream LLM gateways. This documentation outlines an attack surface where malicious input could influence a model, but provides robust mitigation strategies.
- Ingestion points: External request data enters the system through the
inputandinstructionsfields of thePOST /v1/responsesendpoint, as detailed inSKILL.mdandreference.md. - Boundary markers: The
reference.mdfile explicitly states that the adapted chat body must be stripped of all Responses-specific keys before being passed to the gateway. - Capability inventory: The described system performs network operations (HTTPS) to forward requests to an inference gateway.
- Sanitization: The specification defines a strict mapping and '400 vs strip' logic in
SKILL.mdto ensure that sensitive or unsupported fields are either rejected with an error or safely removed before reaching the LLM.
Audit Metadata