AI Stack Alert Medium CVE-2026-115914 For AI Eng, SOC / IR, AppSec, CISO AI-assisted
vLLM DeepSeek-OCR: single crafted PNG crashes engine (CVE-2026-115914), fixed in 0.32.0
vLLM versions 0.11.1 up to but not including 0.32.0 are affected when serving DeepSeek-OCR family models. A zero-sum guard in the image input path returns None for a valid image whose normalized pixels sum to exactly zero. The resulting assertion in the EngineCore worker shuts down the engine process tree, and the advisory says this is reachable by one unauthenticated chat request under default configuration. The fix is in vLLM 0.32.0.
Why it matters
Any exposed vLLM server serving deepseek-ai/DeepSeek-OCR, DeepSeek-OCR-2, or baidu/Unlimited-OCR can be taken offline by one unauthenticated request. The service stays down until a manual restart.
Do this
- Identify vLLM deployments serving DeepSeek-OCR, DeepSeek-OCR-2, or Unlimited-OCR models on versions 0.11.1 through 0.32.0.
- Upgrade affected vLLM deployments to 0.32.0 or later.
- Until patched, restrict network access to the vLLM API endpoint to trusted clients and place it behind an authenticated gateway.
- Alert on EngineDeadError, the message EngineCore encountered a fatal error, and unexpected container exits on OCR inference hosts.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.40%
Signals
- NoCISA KEV
- 6.5CVSS
- 0.40%EPSS
- 27/100Priority
- 80%Confidence
Evidence
“Affected: vllm >= 0.11.1, < 0.32.0 → fixed in 0.32.0.”
“with no per-request containment, escapes to the busy loop's top-level handler”
Sources
Affected / related
vLLM, deepseek-ai/DeepSeek-OCR, deepseek-ai/DeepSeek-OCR-2, baidu/Unlimited-OCR
Published 2026-10-28 21:15 UTC · updated 2026-11-08 02:42 UTC · AI-assisted summary, checked against quoted sources · How we score