threatdeck
Live 14:45 UTC ·

AI Stack Alert High CVE-2026-115912 For AI Eng, AppSec, SOC / IR, CISO AI-assisted

vLLM < 0.32.0: forged multimodal SHM handle reaches worker pickle.loads() (CVE-2026-115912)

vLLM versions before 0.32.0 have a flaw in the multimodal SHM cache used when --mm-processor-cache-type shm is enabled. The worker treats a client-supplied address and monotonic_id pair as authority to read shared memory, and deserializes bytes objects with pickle.loads(). The advisory reports this is reachable through ordinary public HTTP routes without VLLM_ALLOW_INSECURE_SERIALIZATION=1, and the fix is in 0.32.0.

Why it matters

An authenticated inference caller can reach pickle.loads() in the worker process, which the advisory says permits arbitrary Python execution on affected SHM-cache deployments. The opt-in insecure serialization flag does not block this path.

Do this

  1. Upgrade vLLM to 0.32.0 or later on all inference workers and scale-out deployments.
  2. Until patched, disable --mm-processor-cache-type shm if multimodal inputs are served to untrusted callers.
  3. Restrict access to /v1/chat/completions/render, /v1/chat/completions, and /inference/v1/generate to trusted clients.
  4. Hunt for unexpected worker process behavior and EngineCore health endpoint failures (HTTP 503) following inference requests with multimodal features.
  5. Confirm that VLLM_ALLOW_INSECURE_SERIALIZATION=0 does not mitigate this issue; do not rely on that flag as a control.

Exploit Clock

Exploitation unlikely (for now)

  • EPSS 0.51%

Signals

  • NoCISA KEV
  • 7.5CVSS
  • 0.51%EPSS
  • 37/100Priority
  • 85%Confidence

Evidence

“treats that pair as authority to dereference shared memory in the worker”

GitHub Security Advisories (AI packages)

“This is reachable through ordinary public HTTP routes and does not require `VLLM_ALLOW_INSECURE_SERIALIZATION=1`.”

GitHub Security Advisories (AI packages)

“An ordinary authenticated inference caller can execute arbitrary Python in the worker process on a deployment that uses SHM multimodal caching.”

GitHub Security Advisories (AI packages)

Sources

Affected / related

vllm, pickle.loads() deserialization, SHM handle forgery

Published 2026-10-28 21:15 UTC · updated 2026-11-08 02:43 UTC · AI-assisted summary, checked against quoted sources · How we score

Following

Follow the vendors, products, AI tools, techniques and actors you care about. Matching stories, CVEs and exploited bugs are marked, a Following tab appears in the feed, and returning visits show what changed. Saved only in this browser.

Presets

Get updates

A personal RSS feed of new stories on everything you follow. Paste it into any feed reader, Slack or Teams RSS app.