ai tool
vLLM: security advisories, deprecations and releases
High-throughput engine for serving language models.
- Items, 90 days
- 5
- Critical
- 0
- Last critical
- none
- This week
- 1
What your digest would have held this week
1 item from the last 7 days. Zero sends only the ones that touch your stack.
Security 1
- HIGH Google Cloud Blog · 2 Oct 2026
2,076 AI-stack vulnerabilities disclosed
GTIG tracked 2,076 AI-related CVEs from January 2025 to August 2026, concentrated in orchestration frameworks, inference engines, and web portals.
Deprecations 1
- HIGH vLLM releases · 3 Oct 2026
GPTQ activation ordering removed
g_idx activation ordering support dropped; models using it must be re-quantized.
Releases 3
- HIGH NEW THIS WEEK vLLM releases · 2 Nov 2026
vLLM v0.33.0 released with 927 commits, new models and PyTorch 2.14
Major release adds dozens of new models, upgrades PyTorch to 2.14, drops CUDA 12.9 wheels, introduces EFA image, and includes major speculative decoding and performance improvements.
- INFO vLLM releases · 17 Oct 2026
v0.32.0
vLLM 0.32.0 is released with a macOS build fix for the Rust frontend and inherited listener family handling.
- INFO vLLM releases · 3 Oct 2026
v0.31.0: Add Transformers version upper bound in requirements
vLLM 0.31.0 caps the allowed Transformers version to avoid incompatibilities.
Only hear about vLLM when it touches your stack
Tell Zero what you run. You get one email, and only on the days something above applies to you.