CVE-2026-105753 · GHSA-ph3r-5jfg-f84f
vLLM: Mirrored multimodal IPC caches desync after a rejected request — a later request reusing the same media hash trips a receiver assertion in the engine core
- Risk score
- 51/100
- CVSS
- 6.5 v3.1
- Exploit chance (30 days)
- 0.4%
- Published
- 5 Oct 2026
Open in the live forecast Today's forecast
What to know
- What it is
- vLLM is an inference and serving engine for large language models. Prior to 0.28.0, the default mirrored multimodal LRU cache can commit a media hash in the frontend sender cache during multimodal rendering and before engine admission, while the engine receiver cache never receives the payload if that request is rejec…
- Affected
- vllm (pip) < 0.28.0
- Fix
- Upgrade vllm to 0.28.0 or later.
- Exploitation
- No known exploitation
Story map
At a glance
Vllm: Amber, patch this week. vLLM versions 0.22.0 through 0.23.0 fail to validate stop_token_ids against vocabulary bounds in Rust HTTP and gRPC frontends, allowing out-of-vocabulary token IDs to reach MinTokensLogitsProcessor. Attackers can submit requests with min_tokens greater than zero and out-of-vocabulary stop_token_ids to trigger CUDA ten…
The flaws
- CVE-2026-105753: VLLM: Mirrored multimodal IPC caches desync after a rejected request — a later request reusing the same media hash trips a receiver assertion in the engine core (CVSS 6.5 · EPSS 0%)
- CVE-2026-100652: Improper input validation (CVSS 7.5 · EPSS 0%)
- CVE-2026-105759: VLLM is an inference and serving engine for large language models (CVSS 7.5 · EPSS 0%)
- CVE-2026-103241: Denial of service (CVSS 7.5 · EPSS 1%)
The fix
Upgrade Vllm: 0.28.0 · 0.30.0 or later on your release line
Who is exposed: Vllm
Where it stands
- Public proof of concept
Connects to
- Same weakness: CWE-617 (CWE-617). 59 other warnings in 30 days
- Same product: Earlier vllm warnings. 15 in the last 90 days
Why it is Yellow
- CVSS 6.5 medium
- On your watchlist: PyPI
Affected packages
| Package | Ecosystem | Vulnerable | Fixed in |
|---|---|---|---|
| vllm | pip | < 0.28.0 | 0.28.0 |
References
- https://github.com/advisories/GHSA-ph3r-5jfg-f84f
- https://github.com/vllm-project/vllm/security/advisories/GHSA-ph3r-5jfg-f84f
- https://github.com/vllm-project/vllm/pull/46747
- https://github.com/vllm-project/vllm/pull/51897
- https://github.com/vllm-project/vllm/commit/396204230423b7cc6798300926b8fa30190d26a9
- https://github.com/vllm-project/vllm/releases/tag/v0.28.0
- https://github.com/advisories/GHSA-ph3r-5jfg-f84f
Original description
vLLM is an inference and serving engine for large language models. Prior to 0.28.0, the default mirrored multimodal LRU cache can commit a media hash in the frontend sender cache during multimodal rendering and before engine admission, while the engine receiver cache never receives the payload if that request is rejected. A later request reusing the same media hash causes MultiModalProcessorSenderCache to send no payload and MultiModalReceiverCache to reach an assertion with the message "Expected a cached item," producing a shared-service availability failure. This issue is fixed in version 0.28.0.
Last activity 5 Oct 2026. Patch Forecast updates every hour.