An integer overflow occurs in blockIdx.x * 2 * d within the activation_kernels.cu module of vLLM, leading the act_and_mul_kernel to read memory belonging to a different user’s batched request. This flaw allows a request processed in the same inference batch to obtain a partial or complete copy of another user’s inference output, resulting in confidentiality loss. The weakness is a classic integer overflow (CWE‑190).

The vulnerability affects the vllm-project’s vLLM engine, specifically versions prior to 0.27.0. Users running 0.26.x or earlier are at risk; the issue is corrected in release v0.27.0.

The CVSS score of 5.3 indicates a moderate severity. EPSS is not available, and the vulnerability is not listed in CISA’s KEV catalog. Exposure requires that the attacker submit a request that shares an inference batch with a victim; no special privileges are needed beyond being able to submit a request. Consequently, the risk is primarily limited to environments where multiple users perform inference within the same batch, which is common in shared deployments.