What changed in AI, and what it means for your work.

Microsoft says that as early as September 2026, Storm-0940, a China-state aligned actor, used the open-source agentic penetration testing framework ARTEX to automate attacks on hundreds of organizations across many industries. Microsoft assesses that Storm-0940 was responsible for the recently disclosed ConnectWise supply chain compromise of a multi-tenant application, and that several of the ARTEX-driven attacks reached privileged cloud access in less than three hours. Microsoft says follow-on activity through Microsoft Graph included reconnaissance, application credential changes and attempts to modify privileges, and that the activity did not result from a compromise of Microsoft's own systems. It advises customers to review service principal sign-ins, investigate unexpected application credential or privilege changes, and follow ConnectWise's published containment guidance.

Why it matters: If you use ConnectWise or other multi-tenant admin apps linked to Microsoft 365, check for unexpected app credentials and Graph activity now; agentic attack tools can move from first contact to privileged access in hours.

Microsoft AI says its diagnostic system MAI-DxO, published in Science, reached 83% diagnostic accuracy on 304 complex New England Journal of Medicine cases, while physicians scored 22% without external references and 36% with them. On more typical emergency-department cases from MIMIC-IV-ED, Microsoft says MAI-DxO reached 80% accuracy versus 47% for physicians. Rather than one model answering at once, MAI-DxO coordinates a panel of virtual physicians over several rounds, choosing questions and tests and weighing cost; Microsoft says one cost-focused setup kept 81% accuracy at less than half the cost of comparative systems. It is open-sourcing SDBench for select NEJM cases so others can reproduce the evaluation, and it plans prospective testing with Beth Israel Deaconess Medical Center. These figures are from Microsoft's paper.

Why it matters: If you build multi-step agents for any expert workflow, the pattern of several specialised agents debating and picking the next test is more transferable than the medical scores themselves; treat the percentages as one lab's result until others reproduce them.

Microsoft security chief Hayete Gallot wrote on 19 October that "frontier models, regardless of their origin, cannot be trusted". Her reason is not malice: models are trained to get results, which often leads to reward hacking, and agents with broad access may not behave as intended. Microsoft is extending Project Perception, the agent security system it announced in July, to agents running inside a company. It discovers agents, gives each an identity and permissions, scores the risk of what each one is trying to do, and has Guardian Agents supervise their actions in real time. Gallot also says security tools will need to see a model's reasoning, not only its inputs and outputs. This is a vendor announcement, and the post shares no test results or customer data.

Why it matters: If staff use coding or office agents with access to company data, treat each one like a new hire: its own account, only the access it needs, and a log of what it did. That applies whichever vendor you use.

Microsoft announced agent accounts in Copilot Studio, which give an agent its own credentials in Microsoft 365 so people can @mention it in Word, Excel and PowerPoint comments or hand it tasks in Teams and Outlook. Actions are logged to the agent rather than to the person who built it, and admins can scope its access. Business process memory lets an agent carry context from one cycle of a recurring process to the next, and triggers let agents start from an email or a schedule. The features are rolling out through Microsoft's Frontier early access programme, not as a general release.

Why it matters: If your team runs recurring processes such as invoice checks or ticket triage in Microsoft 365, this makes it easier to give an agent a clear, auditable role instead of borrowing a person's login.

Microsoft released Microsoft-Decision-1 on 9 October 2026, available now in Microsoft Foundry and through OpenRouter. It is a "decision model": instead of writing text, it takes a fixed set of options and returns a probability for each, for jobs like routing requests, classifying feedback, approving or flagging agent actions and grading AI answers. It is built on Qwen3.5-9B. Input costs $0.042 per million tokens and output is free. Microsoft says it had the highest accuracy in its own 36-benchmark comparison, ran 2.5 times faster than the runner-up, and changed its answer on only 1.3% of reworded or reordered requests. Internally, it says Xbox Research used it to sort more than 10,000 pieces of player feedback into themes at quality competitive with GPT-6 Sol, over 14 times faster and 200 times cheaper. These are Microsoft's own tests. OpenAI's similar Decisions API entered public beta on 6 October.

Why it matters: If you pay a large model to tag tickets, route messages or approve simple steps, a decision model can do it for a fraction of the cost. Test it on a few hundred of your own labelled examples before switching.