AI Agents — Latest News and Frameworks

Cat Wu will lead product management for OpenAI’s Core Products group after heading product for Anthropic’s Claude Code and Cowork.

Mistral's new Mistral Large 4 Ai model gives organizations control over running and customizing the model for coding and cybersecurity work on their own systems.

GitHub is separating code storage from systems to deliver saved code, aiming to keep extra read traffic from slowing code updates.

OpenAI is rolling out phone messaging support for Dot, its always on task agent, so ChatGPT Pro users can send requests and get progress updates in their phone’s messaging app.

Microsoft generated $18.6 billion in free cash flow in July–September 2026 as stronger cloud collections helped fund AI infrastructure and some spending shifted to later quarters.

Generate Biomedicines used Anthropic's Claude AI to help redesign receptors on immune cells which kept killing cancer cells longer than usual in lab tests.

Apple is adding AI helpers to Creator Studio that can be used to revise documents, slides and spreadsheets, coming in November for subscribers.

Agent Skills lets teams reuse instructions across compatible AI agents. The format has joined the Agentic AI Foundation, with shared governance planned.

OpenAI’s GPT-6.1 Terra model improves coding and computer work at lower prices for developers integrating it into software.

Cohere says network isolation blocked Alibaba's Qwen3.8-Flash-Next to process an unrequested upload of a test string during a cybersecurity task.

DeepSeek's V4.1 Flash 1017 update improves on coding scores and lowers task costs, while factual reliability still trails Claude Opus 5.5.

OpenAI says it cut live internet access for its most advanced AI models during training, replacing it with a virtual web environment where they can still practice browsing.

Harm warnings did not reliably stop Mythos 5 AI’s simulated attacks. Newer Anthropic models showed less harmful behavior, but the cause remains disputed.

Password manager1Password is rolling out user-approved website logins for more AI agents. It can apply previously saved user passwords without giving them to agents directly.

Giving an AI assistant bank login details can mean sharing financial data with an outside provider. Banks want to keep control over whether agents can access their accounts.

The US Army is working on AI agents for cyberwarfare that can move through networks autonomously to infiltrate target systems, posing risks for civilian systems that could be affected.

Meta’s personal AI assistant Muse cannot shop on Amazon. JPMorgan restricts outside agents using bank credentials to move around funds of their customers.

Four Google Gemini 4 Argon AI agents were able to defeat Minecraft's final boss Ender Dragon, writing 2,783 programs and sharing information between each other.

MiniMax's M3.1 Flash AI model improves on M3 in reasoning tests. Across broader AI capability tests, M3.1 Flash's weighted average cost is $0.22 per task, versus M3's $0.51.

AI agents working for a hacker reportedly spent hours breaking into more than 100 company websites.

World ID's new finance suite is designed to help financial firms verify people in video calls and behind AI-agent requests. Its agent toolkit remains in beta.

Anthropic and OpenAI are reportedly negotiating what outside AI safety reviewers can publish. The terms could shape what the public learns about AI risks.

Google is expanding access to Gemini 4 Argon to paying developers, businesses and higher-tier subscribers.

Meta’s new Muse Wonder AI model achieves 68 on the Artificial Analysis Coding Agents Index at max effort in Muse Code, alongside Claude Code with Sonnet 5.5 (max, 68) and Opus 5.5 (max, 66). Muse Wonder (max) scores 63.

Reflection has unveiled its Beam AI model for coding and tool tasks. Downloadable weights to run and adapt Beam are promised for later in October.

Apple plans more explicit consent for Full Disk Access, the Mac permission that lets apps reach private files and messages.

OpenAI has published 722 manuscripts of proposed mathematical solutions from an unreleased AI model, giving researchers new work to examine.

Anthropic's new Claude Haiku 5.5 small AI model adds adjustable reasoning for summaries and delegated tasks and beats OpenAI's cheap GPT-6 Luna model in benchmarks.

Meta has deployed new AI tools to detect Facebook and Instagram ads that lead people to child sexual abuse material elsewhere online.

Claude for Government’s desktop AI assistant is generally available to eligible US federal and state agencies, helping staff work with files and develop software.