vLLM image loader exposes hosts to Ghostscript RCE through unconstrained format dispatch
A high-severity advisory published on October 28 discloses that vLLM's image-loading path can be weaponized to execute arbitrary commands on hosts running a vulnerable Ghostscript build.
vLLM Phi 4 audio encoder OOM bug patched in 0.32.0
A medium-severity vulnerability in vLLM's Phi 4 multimodal audio encoder allowed a single max-length audio input to crash the serving engine through out-of-memory exhaustion. The fix shipped in vLLM 0.32.0; all prior versions are affected.
Critical MLflow RCE gives root on unauthenticated servers
A critical remote-code-execution vulnerability in MLflow gives unauthenticated attackers full control of servers running the open-source ML platform without authentication, according to a GitHub security advisory published on October 27.
ALL BRIEFS
RANKED · 8 THIS DAYONNX Patches Denial-of-Service Flaw in Sub-Byte Tensor Handling
ONNX disclosed CVE-2026-115922, a medium-severity denial-of-service vulnerability affecting all versions through 1.21.0, patched in version 1.22.0.
ONNX Patches Out-of-Bounds Read in ConvTranspose Shape Inference
ONNX disclosed CVE-2026-115920, a medium-severity out-of-bounds read in ConvTranspose shape inference triggered when a weight tensor has fewer than three dimensions and no kernel_shape attribute is provided.
PyTorch 2.15 Ships Rewritten NCCL Backend, Blackwell NVFP4 Decoding, Apple Silicon FP8
PyTorch 2.15, released October 28, makes a rewritten NCCL backend the default for every process group, adding a flight recorder that dumps on hangs and rollback paths via TORCH_DIST_USE_NCCL2=0 or nccl-legacy.
Meta Andromeda Scores 1B+ Ads per Second, 100x Prior Throughput
Meta's next-generation Andromeda ads retrieval engine processes over one billion ads per second, a 100x throughput gain, and has lifted Facebook conversions 2.4% since 2025.
Red Hat Details DRA for Distributed LLM Inference With RDMA and SR-IOV
Red Hat published a technical article on October 28 describing how Dynamic Resource Allocation (DRA) can improve performance of distributed large language model inference workloads using RDMA and SR-IOV.