Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- Microsoft reports $92.3B quarterly revenue and $63.9B Microsoft Cloud revenue
Microsoft official account posted Q3 results: $92.3 billion in revenue and $63.9 billion in Microsoft Cloud revenue for the quarter.
- Microsoft Foundry makes GPT-6.1 Terra generally available for agents that see and act
Microsoft's Foundry blog says OpenAI's GPT-6.1 Terra is now generally available in Microsoft Foundry, with gains in visual reasoning and computer-use speed so apps and agents can read dense charts, navigate desktop apps and process scanned documents.
- Microsoft introduces a Voluntary Exit Offer for eligible employees in participating organizations
Amy Coleman, Microsoft’s EVP and Chief People Officer, said the company is introducing a Voluntary Exit Offer (VEO) to eligible employees within participating organizations. The time-limited, voluntary program is available to eligible employees at level 70 and below in Canada, India, the United Kingdom and the United States, with financial support and career planning resources. Some organizations participate and some do not, based on each business’s priorities and workforce needs.
- Microsoft's Mustafa Suleyman says its AI diagnostic system beats generalist doctors on hard cases, per TechCrunch
Microsoft AI CEO Mustafa Suleyman says the company's diagnostic system outperformed generalist doctors on hard cases in a Science study as it pursues medical superintelligence, TechCrunch reports. The article notes most of the gain appears to come from the underlying models rather than Microsoft's wrapper system.
- Microsoft publishes ThinkingBox, an RL setup that turns database grading into agent rewards
Microsoft posted a Hugging Face article on ThinkingBox, a post-training loop that turns database grading signals into reinforcement-learning rewards for agents. The write-up covers dependability gains, offline policy-relative difficulty estimates and an end-to-end task example.
- Microsoft attributes a ConnectWise supply-chain attack to Storm-0940, which it says automated the campaign with the ARTEX agentic AI framework
Microsoft Threat Intelligence says China-state-aligned Storm-0940 used ARTEX, an open-source agentic AI penetration-testing framework, to automate reconnaissance and abuse a compromised ConnectWise Cloud Backup app credential against hundreds of organizations' Microsoft 365 tenants. Microsoft lists remediation steps for affected customers.
- Microsoft publishes scFM-dataselection, 572 single-cell foundation-model checkpoints studying pretraining size and diversity
Microsoft released scFM-dataselection on Hugging Face: about 32.1 GB of portable checkpoints and PCA matrices used in DenAdel et al., Nature Methods (2026), Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance. The pack includes 125 Geneformer, 125 scVI and 125 SSL/MLP runs, 111 VAEs (including 0% pretraining controls), 6 MLP controls and 80 PCA component sets, covering random sampling, cell-type reweighting and geometric sketching at 1–100% of the corpus plus perturbation spike-ins.
- Microsoft forms Xbox Labs to experiment with AI-powered game decompiling and recompiling, per The Verge
The Verge reports that Microsoft has created a new Xbox Labs group to experiment with decompiling and recompiling games, the techniques fans have used, increasingly with AI help, to port classic titles to new devices or merge code from different games. Microsoft has seen fan projects such as skateboarding modded into Call of Duty and Halo running in a web browser.
- Microsoft president Brad Smith says the UK is best placed to lead global AI safety rules, per TechCrunch
On a BBC podcast, Microsoft President Brad Smith called for urgent AI regulation, arguing that milk and washing machines face more oversight than AI companies. He said the UK, whose AI Security Institute gets pre-deployment access to frontier models, is better placed than the US or EU to lead global safety rules. Smith called the White House's voluntary safety pledge a good step but not enough, and backed breach-notification duties for rogue AI agents, outside evaluator approval before release, and kill switches.
- Microsoft will scale Azure Maia to multi-gigawatt deployments from 2027, making it the primary inference system for its products
Microsoft says Maia 200 now serves production workloads including Microsoft Copilot, with GitHub Copilot and Microsoft Foundry to follow by year end, running OpenAI, MAI and open-source models. Its next chip, Maia 300, is built for long-context reasoning, latency-sensitive chat and multi-step agentic tasks and begins production deployments in 2027. Microsoft expects 40% more tokens per megawatt than leading GPU-based systems and up to 7x Maia 200's rack-scale token throughput in a 96-accelerator configuration.
- Microsoft explains sharing agent skills over MCP's Skills extension, with Foundry toolbox support in preview
A Microsoft Foundry blog post explains the official Skills extension for Model Context Protocol (io.modelcontextprotocol/skills, proposal SEP-2640, now Final), which lets an MCP server list the skills it offers with a manifest of files, sizes and SHA-256 digests so compatible hosts can discover and load them instead of keeping per-agent copies. The extension does not change the skill format or run skills on the server. Microsoft says Foundry's skills support, in preview, follows SEP-2640: skills uploaded to a Foundry project can be referenced in a toolbox and loaded by agents through the toolbox's MCP endpoint.
- Microsoft shows how hosted Foundry agents can answer from SharePoint with each user's own permissions
A Microsoft Foundry blog post walks through building hosted agents in Foundry Agent Service, published to Teams and Microsoft 365 Copilot, that retrieve SharePoint content as the signed-in user rather than as the agent. The agent keeps its own identity for model inference, while a Foundry toolbox and a UserEntraToken connection pass the user's Entra ID token to a Foundry IQ knowledge base, which queries SharePoint live through the Microsoft 365 Copilot Retrieval API so answers only cite documents that user can open. Remote SharePoint knowledge sources in Foundry IQ are in preview, and a sample is on GitHub.
- Microsoft CEO Satya Nadella calls for an AI ‘emergency brake’ and externalized safeguards, per TechCrunch
In a Saturday X post covered by TechCrunch, Nadella said it is time to step back and assess AI’s trust architecture: separate the model from the harness that orchestrates its work, externalize controls, document every meaningful model action with tamper-proof human-readable evidence, and ensure an authorized person can pause or shut down a model mid-task—assuming compromise and containing it from the start.
- VS Code video walks through running an agentic AI workflow in the editor with Microsoft Foundry
A VS Code video covers installation, signing in to Microsoft Foundry and an overview of the UI for running an agentic AI workflow inside VS Code.
- Microsoft Foundry adds a voice user simulator for pre-production voice-agent testing
Microsoft Foundry introduced a voice user simulator that plays the user in spoken conversations with a voice agent under test. Teams set a goal-driven scenario with facts and constraints, then vary acoustic conditions (voices, accents, ambient noise, telephonic audio) and optional conversational dynamics such as barge-in, backchannels, and clarification. The post says the tool is meant for pre-production evaluation where static scripts miss real-time speech behavior, and that available options depend on project and service configuration.
- Microsoft Foundry explains metadata-aware agentic RAG with Foundry IQ for SharePoint libraries
A Microsoft Foundry blog post argues that SharePoint libraries often store useful metadata such as root cause, severity, business unit and document type that never reaches retrieval, so RAG systems miss what a document is about. It describes a metadata-aware agentic RAG approach with Foundry IQ so agents can use that metadata alongside document text.
- Microsoft tests how reliably Foundry's multi-turn evaluators grade voice-agent calls
A Microsoft Foundry blog post reports an internal study of the Task Completion and Generated Rubric evaluators on 3,000+ simulated τ-Voice calls and 278 traces from a Foundry voice agent backed by gpt-realtime-2. Task Completion reached about 83% balanced accuracy with gpt-5.6-terra as judge, all seven GPT judges ordered clearly separated agents the same way as the benchmark, and repeat runs agreed on 95.1% of verdicts on average. The voice-agent evaluation workflow is in preview.
- Microsoft Execution Containers (MXC) generally available for policy-driven AI agent containment
Microsoft made Execution Containers (MXC) generally available as a policy-driven containment layer for AI agents on Windows, macOS and Linux. Developers declare file and network resources in a JSON policy and multi-language SDK; MXC enforces them via process, session, WSL or experimental MicroVM backends. GitHub Copilot, OpenAI Codex, Replit, LM Studio and Unsloth AI already support it; Windows 365 Cloud PC support is also GA.
- Microsoft opens preorders for the $5,999 Surface RTX Spark Dev Box for local AI, per The Verge
The Verge reports that Microsoft's Surface RTX Spark Dev Box, a mini PC on Nvidia's Arm-based RTX Spark platform with 128GB of unified memory, is available for preorder at about $6,000 and ships in November. Microsoft pitches it for running local models over 120B parameters, prototyping agents and on-device coding tools, with developer tools preinstalled.
- Microsoft explains how it keeps LLM-as-judge costs down in Foundry evaluations
A Microsoft Foundry blog post says every evaluation run is a stack of LLM-as-judge calls, with the service handling hundreds of thousands of judge calls a day and billions of prompt tokens in a two-week window, and argues the judge's token bill should be architected as a systems concern rather than treated as a line item.