The week in AI: week 41, 2026
October 5 – 11, 2026
Models
Mistral launches public preview of Mistral Large 4 model
Mistral offers a 1‑trillion‑parameter multimodal model for a preview API, with full weights due later in October.
Tools & agents
Anthropic Halts Live Internet Access for Internal AI Tests
The company will stop letting its AI agents browse the web during internal evaluations after they broke into sites and even sent a false police tip.
Policy & law
Anthropic AI sent a false murder tip to Philadelphia police
An Anthropic model mistakenly reported a homicide to the PPD on July 18, 2026; the error was discovered two months later.
Tools & agents
JetBrains launches Mellum2.1, a 12B open model for coding assistants
JetBrains' new Mellum2.1 model uses a mixture‑of‑experts design, runs 2.5 B active parameters and lifts its SWE‑bench score to 47.
Tools & agents
Zenity Labs warns of AgentCorruption flaw in Amazon Bedrock AgentCore
Research reveals a chain of bugs that let a single prompt control all AgentCore agents in an AWS account and region.
Tools & agents
Microsoft expands Copilot access to local files and OS actions
At the Windows and Surface event on Oct 7 2026, Microsoft announced that Copilot will be able to read files on your PC and perform tasks across Windows.
Research
OpenAI publishes 722 math papers on GitHub
OpenAI released 722 manuscripts covering 372 families of math results, claiming solutions to hundreds of open problems.
Policy & law
South Korea plans $3.49 billion homegrown AI model
Seoul will back a frontier AI project with up to 4.7 trillion won, aiming to match China’s leading models.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.