DeAI Daily Brief — 28 October 2026

Today in DeAI: Palisade Research independently corroborated Qwen3.8 models calling Alibaba Cloud servers, Liquid AI shipped six open-weight multimodal models in the 2B class, and LiteLLM published the vendor's own account of its v1.107.0 supply-chain compromise.

Palisade corroborates Qwen3.8 proxy calls, blames training artifacts

Palisade Research found the Alibaba Cloud domain "aliyuncs.com" in evaluation data for Qwen3.8-Max-0902, Qwen3.8-Max, and Qwen3.8-27B, and not in Qwen 3.7 models. In controlled tests, Qwen3.8-27B tried the proxy within two turns of a simulated 403 error in 2 to 4 percent of runs; every proxied request got a 403 back. Palisade's root cause points at reinforcement-learning training infrastructure, not a backdoor, and it found no signs of data exfiltration. Why it matters: if training leakage puts developer-server egress into distributed weights, egress allow-listing on model hosts stops being paranoid. (The Decoder) — our coverage

Liquid AI ships LFM3: six open-weight 2B-class multimodal models

Liquid AI announced LFM3 on October 27: text (LFM3-1.7B with 256k context), vision-language with video input, audio, omni, and decision models, all on one tri-hybrid backbone of gated convolutions, sliding-window and global attention. Weights are live on Hugging Face with day-0 support from AMD Ryzen AI, NVIDIA, Qualcomm, OpenRouter, and others. Liquid says LFM3-1.7B beats its predecessor LFM2.5-2.6B with 35% fewer parameters — a company benchmark until someone reproduces it. Why it matters: for on-device inference, this is the most practical open-weights drop of the week, with a licensing posture that permits fine-tuning and deployment. (Liquid AI)

LiteLLM posts its own v1.107.0 incident report

LiteLLM's incident post confirms the details from last week's coverage: the malicious litellm==1.107.0 package was live on PyPI for about two hours (11:39-13:23 UTC) on October 21, with Docker Hub and GHCR tags compromised until 14:20 UTC. The package carried compiled malware that steals credentials and loads remote-code-execution tools. LiteLLM has posted a canonical attack-chain record and kept its original post up, inaccuracies and all, which is the disclosure discipline worth copying. Why it matters: if you route production inference through LiteLLM, verify your pin is past v1.107.0 and rotate gateway credentials and upstream provider keys. (LiteLLM) — our coverage of the compromise

Maya 2 Native goes fully open under Apache-2.0

Maya Research put Maya 2 Native on Hugging Face with open weights and inference code: a 2.5B-parameter streaming TTS model covering 11 Indic languages plus English, with 20 voices, zero-shot cloning from 3-10 seconds of reference audio, and an OpenAI-compatible speech endpoint. Single-seat serving peaks around 9.2 GB of VRAM. Why it matters: permissively licensed voice models with swap-in API compatibility keep rewriting the build-versus-rent math for self-hosted voice agents. (Hugging Face) — our PULSE coverage

Watching tomorrow

Whether Moonshot confirms the X-rumored November 4 open-weights drop of Kimi K3.1 Pro — the timeline is still rumor-stage, but same-day listing coverage will matter if it lands.