Skip to content

v1.2.6: two vLLM cross-request advisories

Latest

Choose a tag to compare

@munimv2 munimv2 released this 10 Nov 14:20

Adds two feed entries (210 in total), both for vLLM, both published 9 November 2026 with no CVE ID, and both only for deployments that turn on an optional feature:

  • GHSA-v97p-wpxp-px84 (Medium, CVSS 4.8): only with async scheduling and pipeline parallelism (pipeline_parallel_size > 1) on the V1 model runner, tokens left in a batch slot by an earlier request can be fed into another user's request and end up in its KV cache and output. Affects 0.16.0 up to but not including 0.21.0; fixed in 0.21.0. Workaround on affected versions: --no-async-scheduling or pipeline_parallel_size=1.
  • GHSA-fpch-3xrf-7q96 (Medium, CVSS 5.3): only with reasoning enabled and reasoning_end_str configured, a client that sets the thinking_token_budget sampling parameter can make the reasoning-end string be forced into another concurrent request's output when the batch is reordered. Affects 0.21.0 up to but not including 0.28.0; fixed in 0.28.0.

Neither had an OSV.dev record when checked on 10 November 2026, so most dependency scanners will not report them. The current vLLM release, 0.33.0, is affected by neither. Users still on 0.16 to 0.27 were already told to upgrade by earlier entries (for example GHSA-5r3c-5x4v-8gjh, fixed in 0.33.0). The GitHub Action example now pins v1.2.6. Apache-2.0. Made by MV2 of Munim, Inc., with AI agents.