A DeepSeek-OCR image preprocessor in vLLM treated a zero-valued pixel sum as if the image embeddings were absent. The reproducer used a legitimate image whose normalized black and white pixels canceled out rather than a corrupted upload. The downstream embedding path failed and terminated the EngineCore service instead of returning a request-local error. T…
Inspect the ClaimsCase · DiggingBeagle record
vLLM DeepSeek-OCR rejected a valid balanced image as a missing embedding
A DeepSeek-OCR image preprocessor in vLLM treated a zero-valued pixel sum as if the image embeddings were absent. The downstream embedding path failed and terminated the EngineCore service instead of returning a request-local error.
Scope
The reproducer used a legitimate image whose normalized black and white pixels canceled out rather than a corrupted upload. The condition concerns the affected OCR model path and does not imply that every image model in vLLM is susceptible.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Evidence
- Grade B
- Remediation
- patch available
- AI role
- AGAINST AI
Basis and provenance
The cited primary technical report or advisory supports the stated mechanism within its reported scope. The assessment does not infer a production breach, a victim count or a comparable impact score from a laboratory result. A maintainer or advisory names a patched release; that does not prove every deployed instance has updated. Boundary: Numeric zero is a legitimate result of a sum, not a sentinel indicating absent image content.
Assessed Oct 29, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-CP23-VLLM-DEEPSEEK-OCR-ZERO-1 · CLM-ENRICH-20261029-35 · vLLM DeepSeek-OCR valid zero-sum image engine crash, CVE-2026-115914
At a glance
The downstream embedding path failed and terminated the EngineCore service instead of returning a request-local error.
Read the implicationsThe condition concerns the affected OCR model path and does not imply that every image model in vLLM is susceptible.
Limits and uncertaintyFull account
A DeepSeek-OCR image preprocessor in vLLM treated a zero-valued pixel sum as if the image embeddings were absent.
The reproducer used a legitimate image whose normalized black and white pixels canceled out rather than a corrupted upload. The downstream embedding path failed and terminated the EngineCore service instead of returning a request-local error.
The condition concerns the affected OCR model path and does not imply that every image model in vLLM is susceptible. Upgrade to vLLM 0.32.0 and validate tensor shapes and missingness without equating numeric zero to absent content.
Evidence interpretation and control boundary. Numeric zero is a legitimate result of a sum, not a sentinel indicating absent image content. The affected DeepSeek-OCR preprocessors used a zero-sum tensor condition to skip embedding creation, and a later component treated missing embeddings as a fatal engine error. A request-local validation failure should not propagate into process-wide inference shutdown. This does not imply all image input paths are affected.
Mechanism and trust boundary
- 01
Step
- 02
Step
- 03
Step
Claims & evidence
CLM-ENRICH-20261029-35The vLLM advisory reproduces engine failure using valid DeepSeek-OCR image pixels whose normalized values sum to zero, rather than an invalid file format.supported
Basis: reported finding
- supportsvLLM DeepSeek-OCR valid zero-sum image engine crash, CVE-2026-115914primary disclosure
GHSA-gq5q-89c7-vhvx: normalized pixel sum, DeepSeek-OCR reproduction and fix
CLM-CP23-VLLM-DEEPSEEK-OCR-ZERO-1A DeepSeek-OCR image preprocessor in vLLM treated a zero-valued pixel sum as if the image embeddings were absent. The reproducer used a legitimate image whose normalized black and white pixels canceled out rather than a corrupted upload. The downstream embedding path failed and terminated the EngineCore service instead of returning a request-local error.supported
Basis: reported finding
- supportsvLLM DeepSeek-OCR valid zero-sum image engine crash, CVE-2026-115914primary disclosure
GHSA-gq5q-89c7-vhvx: normalized pixel sum, DeepSeek-OCR reproduction and fix
Implications
The downstream embedding path failed and terminated the EngineCore service instead of returning a request-local error.
Controls and mitigations
- Upgrade to vLLM 0.32.0 and validate tensor shapes and missingness without equating numeric zero to absent content.
Unknowns and contradictions
- The condition concerns the affected OCR model path and does not imply that every image model in vLLM is susceptible.
Sources and citation
Material revision history
- Oct 29, 2026 · Canonical change recorded · new in release · revision 202