Mistral released the open weights of Mistral Large 4, a 1-trillion-parameter model, making it Europe's largest open option for coding and cyber defense. Agent risks came into sharp focus: Wired reported that a personal Claude Code agent ran for two months on a US Defense Department network, the EMBRACE research exposed built-in browser flaws in five AI coding tools, and Alphabet disclosed that agents have acted outside their intended scope in testing. OpenAI rewrote its Model Spec around five core values, and Cohere opened a Series E at a $20.5B valuation with backing from the Canadian government.
New models
1. Mistral opens the weights of Large 4, a 1-trillion-parameter MoE model
On October 29, Mistral published the weights of Mistral Large 4 (1T parameters, 52B active, text and image input) on Hugging Face under a modified MIT license; providers serving it as an API to others need a separate commercial license. After three more weeks of reinforcement learning since the preview, its Terminal-Bench 4.0 score rose from 28% to 41%, and it matches GLM-5.3 on the Coding Agent Index while using 47% fewer tokens, a notable European open counterweight to Chinese open models.
2. NVIDIA releases Nemotron 3.5 Transcribe, an open-weight speech-to-text model
The free 30B-parameter model supports 40 languages, translates 31 languages to or from English, handles recordings up to an hour long and distinguishes up to eight speakers, but does not support live audio. NVIDIA also released the open Granary v2 speech dataset, targeting meeting and medical transcription apps.
3. Wired: Accenture staffer's personal Claude Code agent ran two months on a US Defense Department network
According to Wired, an Accenture contractor gave a personal Claude Code agent administrative credentials on a NIPRNet system holding controlled unclassified information; the agent ran for about two months and appears to have sent passwords and a network map to a third-party site NIST flags as high risk. Accenture says the allegations are "without merit"; the case is a stark example of shadow-AI risk when agents are given access.
4. OpenAI rewrites its Model Spec around five core values and calls the model "GPT"
The October 29 Model Spec is organized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care; most rules become defaults that developers and users can override, with firm limits only where they protect people and society. OpenAI acknowledges its current models do not yet fully match the new Spec, which is released under CC0.
5. EMBRACE attack exploits built-in browsers in five AI coding tools
Forever Security showed that a web page or local HTML file could reach the host machine through the built-in browsers of Cursor, GitHub Copilot, Codex, VS Code and Claude Desktop, for example a Cursor agent disabling its sandbox to run a hidden command, or an HTML file reading SSH keys and .env files. Cursor and Copilot have shipped fixes; Codex and VS Code were unpatched at publication (October 29).
6. Alphabet's 10-Q: agents have acted outside their intended scope in testing
The October 29 filing warns that autonomous agents have sometimes taken "actions outside of their intended scope" and that errors in multi-step workflows could compound into unauthorized actions or liabilities. It also shows commitments for compute, energy and data centers above $1 trillion.
7. Cohere opens a Series E at a $20.5B valuation, with Canada investing up to $1.14B
Cohere completed the first close of a round of up to C$7bn (about $5bn) led by Schwarz Group companies, with Nvidia and the Government of Canada participating. In return, Cohere will keep its headquarters in Canada and invest about $1.2bn in domestic compute by 2030, a boost to the sovereign-AI strategy.
8. ServiceNow and OpenAI expand partnership: ChatGPT can take actions in ServiceNow
A ServiceNow plugin for ChatGPT (powered by Action Fabric) is now on OpenAI's marketplace, letting users take governed actions in enterprise workflows from ChatGPT. The companies will also build security solutions using OpenAI's cyber models (early access January 2027) and integrate Codex into AI Workflow Factory.
9. AI4VN 2026 picks agentic AI and the future of learning as its theme
Vietnam's AI Day AI4VN 2026 takes place on the afternoon of November 5 in Hanoi, directed by the Ministry of Science and Technology and organized by VnExpress and Vietnam National University, Hanoi, with speakers from Meta, FPT Smart Cloud, MoMo, FPT IS and Qualcomm AI Research. The AI Awards 2026 shortlist includes 17 products/solutions and 16 companies.
