vLLM Derender Endpoints Lack Output Bounds, Pose Resource Risks

According to the GitHub Security Advisory database, a medium-severity issue has been identified in vLLM's `/v1/completions/derender` and `/v1/chat/completions/derender` endpoints. These endpoints decode caller-supplied `GenerateResponse` token IDs without enforcing output bounds such as model context length, `max_tokens`, choice count, or response size. This allows authenticated API clients to supply unbounded generated-output-shaped JSON, potentially causing excessive CPU and memory usage on the server.

Operators should verify whether their vLLM deployment exposes these derender endpoints and confirm that API key authentication is enforced. If these endpoints are not essential for your use case, consider disabling them or restricting their usage to trusted clients. Additionally, evaluate whether your infrastructure can handle potential resource exhaustion scenarios caused by unbounded input data. This issue highlights the importance of validating input and enforcing strict bounds, a common pattern in secure API design.