Evidence · DiggingBeagle record
vLLM advisory GHSA-ph72-cqr5-qpp7: untrusted scale-out features
September 23 GitHub advisory for disaggregated multimodal rendering/generation trust boundary: generation accepts caller-supplied internal features and can cause engine DoS or cache corruption.
- Published
- Sep 23, 2026
- Publisher
- vLLM security advisory
- Source role
- primary disclosure
Evidence record
September 23 GitHub advisory for disaggregated multimodal rendering/generation trust boundary: generation accepts caller-supplied internal features and can cause engine DoS or cache corruption.
Claim-level citations (2)
- supportsvLLM scale-out multimodal generation trusts forged render state: The advisory states POST /inference/v1/generate trusts caller-supplied serialized encoder tensors and multimodal metadata without sufficient binding to a trusted render result.
Summary: render and generate routes
- supportsvLLM scale-out multimodal generation trusts forged render state: The advisory lists affected versions below 0.30.0 and a patch in 0.30.0.
Package: affected and patched versions