Use of Uninitialized Resource Affecting vllm package, versions [0.8.5,0.31.0)
Threat Intelligence
The probability is the direct output of the EPSS model, and conveys an overall sense of the threat of exploitation in the wild. The percentile measures the EPSS probability relative to all known EPSS scores. Note: This data is updated daily, relying on the latest available EPSS model version. Check out the EPSS documentation for more details.
Do your applications use this vulnerable package?
In a few clicks we can analyze your entire application and see what components are vulnerable in your application, and suggest you quick fixes.
Test your applications- Snyk IDSNYK-PYTHON-VLLM-21707814
- published25 Oct 2026
- disclosed25 Oct 2026
- creditLuis Chamberlain
Introduced: 25 Oct 2026
NewCVE-2026-116180 (opens in a new tab)How to fix?
Upgrade vllm to version 0.31.0 or higher.
Overview
vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs
Affected versions of this package are vulnerable to Use of Uninitialized Resource via LogprobsTensors.empty_cpu() in vllm/v1/outputs.py, which allocates an uninitialized int32 tensor using torch.empty(). When the V1 model runner processes a request whose prefix is reported as externally computed by a KVConnectorBase_V1 subclass, positions [0 .. num_computed_tokens - 1] are never written into the tensor. A client sharing the engine can send a request with prompt_logprobs set (or logprobs with echo=true) against a connector-cached prefix, causing tokenizer.decode() to encounter garbage int32 values and raise OverflowError, which crashes EngineCore and drops all co-tenant in-flight requests.
Note: This is only exploitable when the V1 engine is configured with a KVConnectorBase_V1 subclass (e.g. LMCacheConnectorV1, NixlConnector) whose get_num_new_matched_tokens() returns a value greater than zero.
Workaround
Disable KV connectors on the affected engine by removing --kv-transfer-config and any kv_transfer_config setting. Alternatively, strip prompt_logprobs, logprobs, and echo from inbound requests at an API gateway or middleware layer in front of the engine.