vLLM versions before 0.28.0 fail to validate the lower bound of token IDs in the /v1/embeddings and /pooling endpoints, allowing unauthenticated attackers to crash the engine by submitting negative token IDs. A single request with a negative token ID triggers a CUDA device-side assertion that poisons the GPU context, causing all subsequent requests to fail until the process restarts.
Improper neutralization of input during web page generation ('cross-site scripting') in Azure Portal allows an unauthorized attacker to perform spoofing over a network.
vLLM through 0.29.0 fails to properly clean up decode-side metadata for rejected inference requests in prefill/decode disaggregated deployments. Remote attackers can submit requests with max_tokens=0 to exhaust decode-worker memory without bound until the worker restarts.
Improper neutralization of special elements used in a command ('command injection') in M365 Copilot allows an authorized attacker to elevate privileges over a network.
Improper neutralization of special elements in output used by a downstream component ('injection') in Azure Cosmos DB allows an authorized attacker to elevate privileges over a network.