Database/AI/ML frameworks & serving
vLLM: an all-zero image on the DeepSeek-OCR path trips a fatal assertion and shuts the engine down
Impact
On the DeepSeek-OCR, DeepSeek-OCR-2 and Unlimited-OCR image paths, an image whose normalized pixels sum to zero produces no embedding when it is alone in its per-step encoder batch. The empty value propagates through embed_multimodal to sanity_check_mm_encoder_outputs, where a fatal assertion fires and the engine shuts down for all users. The input is a legal image, so nothing upstream rejects it, and a single repeated request is enough to keep a shared OCR endpoint down. Recovery is a process restart with a full weight reload onto the GPU, so the blast radius is every tenant on that server.
Who can reach it
Any remote client that can reach the OpenAI-compatible API and submit an image (CVSS: authenticated, low privileges) against a server running one of the affected DeepSeek-OCR models.
What to do
Upgrade to vLLM 0.32.0 (fix commit c15672cf) and restart the serving replicas - a daemon restart plus model reload, no host maintenance. Until then, reject all-zero or blank images at the ingress, or keep the affected OCR models off publicly reachable endpoints.
References
Related entries
- vLLM: off-by-one prompt-token-ID bound lets a request crash the inference engineCVE-2026-115916 · vLLM tokenizer input validation (vllm/tokenizers/hf.py max_token_id)Medium
- vLLM: request-supplied chat template bypasses trust_request_chat_template and exhausts the serverCVE-2026-115918 · vLLM /v1/responses endpoint (chat_template_kwargs trust policy)Medium
- MLflow: CrewAI autologging writes provider API keys in plaintext into tracesCVE-2026-116691 · MLflow CrewAI autologging (mlflow/crewai/autolog.py agent attribute capture)Medium
- Kubeflow Pipelines: frontend artifact routes skip namespace authorization, exposing other tenants' artifactsCVE-2026-118760 · Kubeflow Pipelines frontend (artifact routes, namespace authorization)Medium
- vLLM: unbounded frame count in video/jpeg base64 data URLs crashes the server with OOMCVE-2026-34755 · vLLM OpenAI-compatible API server (video/jpeg base64 multimodal path)Medium
- vLLM: no upper bound on the n parameter lets a single request OOM the API serverCVE-2026-34756 · vLLM OpenAI-compatible API server (ChatCompletionRequest/CompletionRequest n parameter)Medium
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.