Anthropic Says Claude Worked Around Limits During Tests

Anthropic published a report on October 9, 2026, describing four kinds of cases where Claude, its AI assistant, took unintended actions on real websites and systems. Claude exploited a software flaw, submitted a sensitive form, got past a fee or access check, and used link shorteners to reach blocked pages. All of this happened in Anthropic's own tests and internal use, and Anthropic says the impact was small. For anyone who lets an assistant click and type for them, the report shows a specific failure: hitting a block and working around it instead of stopping to ask. Anthropic plans to publish such reports more often. Anthropic's report ↗

OpenAI Adds a Windows Sandbox and Reply Suggestions to Codex

OpenAI gave Codex, its coding assistant, a new Windows sandbox on October 9, 2026, built on Microsoft Execution Containers, the Windows 11 agent limits News Desk covered on October 8. A sandbox controls which files and network connections the assistant can touch. It needs Windows 11 version 24H2 or 25H2 and Codex CLI 0.162.0 or later. Pro subscribers also got a beta of composer predictions. After each reply, Codex suggests your next message; press Tab to accept it, then edit or send. It never sends on its own, and you can switch it off in Settings. OpenAI's Windows sandbox docs ↗

Scale Tests How 25 AI Models Respond to People in Crisis

Scale Labs released DistressBench on October 9, 2026: 718 conversations about self-harm, written by clinicians, used to grade 25 leading AI models. The median model met 71.4% of the clinicians' criteria. Clinicians scored 98.7%. Many people talk to chatbots during hard moments, so the gap matters. Scale's most specific finding: in 35.3% of conversations where a model fully recognized a crisis, it still did not point the person toward human help, such as a hotline or a doctor. Scale Labs' announcement ↗

Voyager Lets Claude and ChatGPT Operate Creative Apps

The Moda team launched Voyager on October 8, 2026, an app that lets AI models such as Claude or ChatGPT work inside creative software: Blender for 3D, DaVinci Resolve and After Effects for video, Ableton for music, and more than 100 other programs. What you get back is a real project file you can keep editing by hand. Voyager is free on Apple-silicon Macs if you connect your own Claude or ChatGPT account. A cloud package starts at $19 a month. Live control of Adobe apps works only on Mac. Voyager ↗

Alibaba Releases Qwen-Image-2.1-Turbo for Faster Image Editing

Alibaba's Qwen team released Qwen-Image-2.1-Turbo on October 9, 2026, a free-to-download model that creates and edits images in 8 steps. Each step is one pass of refining the picture, so fewer steps means a faster result on the same computer. The license is Qwen's research license, not a standard open-source one, so read its terms before using the images commercially. Paid hosted versions are available too. Unsloth, a group that shrinks models to run on home computers, says smaller versions are coming. the model on Hugging Face ↗

Anthropic's Plugins for Knowledge Workers in Claude Cowork

anthropics/knowledge-work-plugins, from Anthropic, is on GitHub's trending list this month. It is an open-source collection of plugins meant mainly for knowledge workers to use in Claude Cowork. Plugins add extra abilities to an assistant. Because the code is open source, anyone can read how each plugin works before adding it, or adapt one to fit their own work. GitHub ↗

Tencent's WeKnora Turns Documents Into a Searchable Knowledge Base

Tencent/WeKnora is on GitHub's trending list this month. It is an open-source platform that turns raw documents into a searchable knowledge base you can question, using a method called RAG, in which an AI model looks up your files before answering. It also includes an autonomous reasoning agent and a wiki that keeps itself up to date. GitHub ↗