AI beats GPs on hard cases, AI agents hit a supply chain, and Amex cuts 5% for AI

Microsoft's AI unit has published its diagnostic system, MAI-DxO, in the journal Science. The paper, "Sequential diagnosis with language models", appeared on 22 October. MAI-DxO is an orchestrator: it makes several AI agents behave like a panel of doctors with different roles, deciding what to ask a patient, which tests to order and when to commit to a diagnosis. On 304 difficult case records from the New England Journal of Medicine, turned into step-by-step simulated encounters, the system reached the correct diagnosis about 83% of the time. Primary care doctors and general internists working alone with reference materials got about 36% right, according to TechCrunch and other coverage of the paper.

The study also tested 325 emergency department cases from Beth Israel Deaconess Medical Center in Boston, where reports put MAI-DxO at about 80% against roughly 47% for physicians. The benchmark, called SDBench, gives the participant a short summary and makes them pay, in estimated dollars, for every question and test. The best results came when MAI-DxO ran on OpenAI's GPT-5 and GPT-5.5. With GPT-5.5, complex-case accuracy rose from about 72% for the model alone to nearly 87% with the orchestrator, at slightly lower estimated cost. A weaker model, Gemini 2.5 Flash, jumped from 46.4% to nearly 69%.

This is the peer-reviewed version of work Microsoft first released as a preprint in June 2025, when it claimed up to 85% accuracy. The figures have changed, so the Science numbers are the ones to use. Microsoft is not alone: Google and OpenAI have also claimed their systems can match or beat doctors on some tasks, and Google's AMIE was separately studied with Beth Israel in real urgent-care visits. Mustafa Suleyman, who runs Microsoft AI, called the result a step towards "medical superintelligence" and said he expects healthcare models to get an order of magnitude better within 24 to 36 months. Microsoft is also building a healthcare foundation model with the Mayo Clinic.

The caveats are large. These were simulations built from published past cases, not live patients with vague symptoms and missing records. Complex cases like these are normally handled by specialists or teams, who were not compared. All 15 authors work for Microsoft. And the paper's own data suggests most of the improvement comes from the stronger underlying model, not the orchestration: on emergency cases, GPT-5.5 scored about 85% with or without MAI-DxO. Cost savings are estimates from a uniform price list, not measured hospital spending, and one independent analysis found the emergency-case cost was higher than the physicians'. No product or release date has been announced.

The practical takeaway is that structured, step-by-step clinical reasoning by AI is now strong on paper, and the next phase is proof in real hospitals. Beth Israel will run a study in which consenting patients receive normal care and also interact with MAI-DxO, so its recommendations can be compared with clinicians'. For health systems in places with few specialists, including much of Africa, the promise is obvious, but so is the risk of deploying tools validated on US case records. Until local trials and regulators catch up, the safest near-term uses are in training, case review and decision support where a qualified clinician stays in charge.

Why it mattersIt is the strongest peer-reviewed evidence yet that AI can reason through hard diagnoses step by step while tracking test costs, and it shows the base model matters more than the wrapper around it.

Microsoft published a report on 21 October attributing the recent compromise of ConnectWise's Cloud Backup service to Storm-0940, a China state-aligned group active since 2021. What makes the case stand out is the tool. Microsoft says the group ran its operation through ARTEX, an open-source penetration-testing framework in which multiple AI agents handle reconnaissance, password attacks and privilege escalation with little human input. Microsoft says ARTEX let the attackers "accelerate the activity to machine speed". In multiple cases, less than three hours passed between first targeting an organisation and gaining privileged access to its cloud infrastructure.

The numbers in Microsoft's report are specific. Between 10 and 14 October, infrastructure linked to the group ran an automated password-spray campaign against 327 accounts at eight organisations, generating more than 151,000 failed logons and peaking at 2,900 failures a minute on 11 October; Microsoft saw no successful logins from that spray. The real damage came through ConnectWise. The attackers gained control of the identities of several ConnectWise cloud applications, which customers had already trusted inside their Microsoft 365 tenants, then tried to sign in to more than 2,250 tenants. One chart shows over 200 sign-ins in under four minutes using four app credentials.

Because the attackers logged in as trusted applications rather than as people, no interactive user sign-in was needed, so the usual user protections such as login prompts were not in the path. Inside some tenants, Microsoft says, the agents ran reconnaissance through Microsoft Graph, added new app credentials for persistence, changed groups and privileged roles, read secrets in bulk from Azure Key Vault, and in one case had an app grant itself Global Administrator. ConnectWise published its first advisory on 16 October and suspended several services, including Cloud Backup, according to SecurityWeek. Security firm Huntress reported access to hundreds of tenants.

Some details are still unclear. Reports differ on how many customers were actually compromised, as opposed to targeted: the 2,250 figure is sign-in attempts, and ConnectWise has not given a final count. Coverage also disagrees on whether ARTEX remains openly available. Microsoft's attribution has not been independently confirmed, and China typically denies such claims. Microsoft says multiple threat actors in different regions are using ARTEX, and SecurityWeek notes that CrowdStrike, Zscaler and Recorded Future had earlier linked similar AI-powered activity to attacks on South Korean financial institutions. Zscaler published its own technical analysis of the tooling.

The lesson for everyone else is that supply-chain apps are the soft spot, and AI makes exploiting them fast. Most small and mid-sized firms have dozens of third-party apps with standing permissions in their Microsoft 365 or Google Workspace accounts, and few ever review them. Microsoft's advice is concrete: keep affected ConnectWise connections disabled, preserve logs before deleting anything, inventory apps by application ID, review service-principal sign-ins back to 1 October, revoke unneeded permissions, rotate credentials and monitor for new credentials or role changes. That is ordinary hygiene, but at machine speed there is no time to do it after the fact.

Why it mattersIt is one of the clearest documented cases of a state-aligned group using an AI agent framework to run a real supply-chain intrusion, and it targeted the trusted third-party apps most businesses never audit.

American Express disclosed on Friday, with its third-quarter results, that it expects to eliminate about 5% of its positions worldwide and take a $336 million pretax restructuring charge. CEO Stephen Squeri put the figure at about 3,500 jobs, according to reports citing Bloomberg; outlets that applied 5% to the nearly 77,000 staff Amex had at the end of 2025 estimate closer to 3,800. Chief financial officer Christophe Le Caillec told CNBC the goal is to make the company "leaner, more agile, and also to prepare for AI." Part of the cost is funded by the gain from selling its stake in Global Business Travel Group.

The explanation is more specific than the usual 'AI took the jobs' framing. Le Caillec said Amex wants to reduce management layers and widen spans of control first, because "you don't want to automate processes that are not optimally structured." In other words, the cuts are groundwork for automation rather than a direct swap of people for software. The results themselves were solid: revenue rose 9% to about $20.1 billion, adjusted earnings per share were $4.68 against analyst expectations of $4.48, and Amex raised its full-year earnings guidance. Gen Z cardholders' US spending grew 35% in the quarter. Shares still fell 2–3% in early trading.

Amex also laid out how it will treat outside AI agents, a growing question for every company that handles payments or bookings. Squeri said the company is "not going to let assistants run rogue through our system" and will only allow agents that come through approved API or MCP channels; MCP is the open protocol many AI assistants use to connect to tools. Agents will not be allowed to repeatedly query its Resy restaurant booking service in ways that disadvantage other cardholders. Amex says it is developing purchase-intent authentication and protections for cases where registered agents make mistakes, and it is working with OpenAI on an Amex Offers plugin.

There are reasons for caution in reading this as pure AI disruption. Squeri presented the cuts as an efficiency move taken from strength and did not attribute the whole reduction to automation, and companies often give several reasons for layoffs at once. The exact job count differs between sources. Agentic payments are still small, Squeri said, and he called it too early to judge how AI will affect costs. Gartner has predicted that half of companies that cut staff because of AI will rehire by 2027, often under different job titles. Amex is also not alone: Barclays confirmed this week that it plans to cut about 5,000 jobs, roughly 5% of staff, in a technology-driven cost programme.

For workers and small firms, two signals matter. First, large employers are now restructuring before they automate, which means demand is shifting towards people who can map, simplify and document processes, and away from layers whose job is mainly passing information along. Second, big merchants are drawing hard lines on how AI agents may interact with them: approved APIs and MCP connections in, scraping and uncontrolled bots out. Any business that wants AI assistants to book, buy or query on its customers' behalf will need a proper, permissioned interface, and many smaller businesses have no idea how to build one.

Why it mattersA highly profitable company is cutting staff explicitly to prepare for automation, and one of the biggest card networks has set a template for how AI agents must connect to merchants.