Snyk has a proof-of-concept or detailed explanation of how to exploit this vulnerability.
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.
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 applicationsUpgrade vllm to version 0.28.0 or higher.
vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs
Affected versions of this package are vulnerable to Improper Validation of Array Index via InputProcessor._validate_model_input in the /v1/embeddings and /pooling endpoints, which validates only the upper bound of caller-supplied token IDs but not the lower bound. An unauthenticated attacker can submit a single request containing a negative token ID (e.g. -1), which bypasses validation and is used directly as an index into the GPU embedding table, triggering a CUDA device-side assertion inside PyTorch's index kernel. This assertion poisons the CUDA context, causing all subsequent CUDA operations in the process to fail, crashing the engine and rendering every subsequent request from every client unserviceable until the process is restarted.
Users that are not able to upgrade to the fixed version are advised to:
Reject request bodies containing negative integers in input at a reverse proxy or API gateway in front of vLLM.
Do not expose the token-id form of the input; accept only string input from untrusted callers.
Restrict network access to /v1/embeddings and /pooling.