CGA-qq6v-fx2w-3hp8

GHSA-q8gq-377p-jq3r (CVE-2026-41523) is a security check bypass via assert in activation functions (High) vulnerability in vLLM. The vulnerable component is vllm itself (0.18.1), not an updatable dependency: upstream's fix first appears in vllm 0.22.0.

The py3-vllm-cuda-12.4 stream is pinned to vllm 0.18.1 because it is the last release supporting CUDA 12.4; vllm 0.19.0 and later require CUDA 12.8 or newer (cuda::ptx::fence_proxy_async). Upstream ships no 0.18.x maintenance release: the releases/v0.18.1 branch is identical to the v0.18.1 tag. No upstream release this stream can consume carries the fix, and per the CVE remediation patch policy a backport of the upstream commit is an exception requiring prior approval, so it is not being carried here.

This advisory will be updated if an upstream release compatible with CUDA 12.4 ships the fix. Customers who need the fix today should migrate to the vllm-openai-cuda-13.0 image stream, which carries vllm 0.28.0 (the fix landed in 0.22.0).

Fixed upstream in vLLM 0.22.0 (commit b3c7ffcab82c2439726f8cb213800f6f38c023d3, PR vllm-project/vllm#43286). This package is pinned to vLLM 0.18.1 for CUDA 12.4 compatibility: vLLM >= 0.19 requires CUDA >= 12.8 to compile, so the fixed release cannot be built on this toolchain. Backporting the fix to 0.18.1 touches a refactored pooling-metadata layout and does not apply cleanly.