vLLM fixes three cross-request cache and state isolation vulnerabilities in a row

GitHub security advisories collected from 2026-10-06 to 10-08 show that vLLM fixed three cross-request isolation defects in succession. First: the flash late-interaction path derives shared query embedding cache keys from the client-controlled X-Request-Id, causing cross-request cache conflicts, incorrect scores, or cache misses in /score and /rerank; only that flash path is affected, and the maintainers switched to a server-generated request namespace to isolate the cache; affected versions are < 0.30.0, fixed version 0.30.0. Second: the multimodal EXIF hash and prefix cache extra keys accept client-controlled values and are not sufficiently bound to the request body and media content, which may cause cross-request cache confusion and break the isolation boundary between media processing and prefix caching; affected versions are < 0.31.0, fixed version 0.31.0. Third: when the V1 engine reuses batch slots, the allowed_token_ids mask from a previous request remains, causing subsequent requests to inherit output vocabulary constraints set by others; even if HTTP 200 is returned, this is vocabulary steering rather than arbitrary text replacement; the advisory does not report confidentiality or availability impact; affected versions are < 0.31.0, fixed version 0.31.0.

Generated by AI from reports · Updated 2 days ago

1. Oct 6, 2026 · 2 reports
   vLLM shared cache key trust boundary cross-request confusion vulnerability
   GitHub 安全公告：vLLM discloses shared cache key trust boundary flaw that can cause cross-request cache confusion
2. Oct 6 07:42 · 1 reports
   vLLM flash late-interaction cache key cross-request contamination vulnerability
   GitHub 安全公告：vLLM fixes cross-request cache key collision vulnerability affecting /score and /rerank

No discussion links have been collected for this event yet.

Hourly attention