CVE-2026-71486

Name of the Vulnerable Software and Affected Versions vLLM versions prior to 0.26.0

Description The '/v1/completions/derender' and '/v1/chat/completions/derender' endpoints accept caller-supplied GenerateResponse objects. The generate responses, choices, token ids, prompt logprobs, logprobs.content, top logprobs, and routed experts structures are processed by OnlineDerenderer and tokenizer.decode before limits such as max model len, max tokens, max num seqs, or response-size are enforced. This allows an authenticated API client to cause excessive CPU and memory consumption and generate oversized responses.

Recommendations Update to version 0.26.0.