Oct 23, 2026

Microsoft AI Tops Doctors on Hard Cases; Hackers Use Agents

Microsoft's MAI-DxO more than doubled generalist doctors' accuracy on hard cases in a Science study, as Microsoft warned that state-aligned hackers used AI agents in a ConnectWise supply-chain attack.

Listen to this update

Microsoft says its multi-agent diagnostic system more than doubled the accuracy of general physicians on hard medical cases, in a study now published in Science. The result led a busy Thursday in tech and AI that also brought a Microsoft warning about state-aligned hackers running AI agents, a large-scale software verification effort from Anthropic, and two very large funding rounds for OpenAI and DeepSeek.

Microsoft's MAI-DxO beats generalists on complex cases

Microsoft AI's system, called MAI-DxO, coordinates several AI agents that behave like a panel of clinicians. They question a simulated patient, order tests and argue over possible diagnoses before settling on an answer. In a simulation built from difficult case reports in the New England Journal of Medicine, the system reached the correct diagnosis 83% of the time. Primary care doctors and general internists working alone, with access to reference materials, got 36% right.

The comparison has clear limits. Specialists and physician teams, who normally handle cases this complex, were not part of the test, and every case was simulated rather than drawn from live patient care. TechCrunch also notes that the study's own data suggests most of the gain comes from the underlying language models, with Microsoft's orchestration layer adding a smaller boost on top. MAI-DxO was tried with models from OpenAI, Google, Anthropic, Meta and DeepSeek, and did best with OpenAI's GPT-5 and GPT-5.5.

The figures show both sides. Paired with GPT-5.5, the system reached nearly 87% accuracy on complex cases, compared with 72% for the model on its own, at slightly lower cost. On emergency department cases, the orchestration made little difference to either accuracy or cost. Microsoft AI chief Mustafa Suleyman described the work as a step toward "medical superintelligence." The company plans to test MAI-DxO against human doctors in real-world settings with Beth Israel Deaconess Medical Center in Boston. It is also building a healthcare foundation model with Mayo Clinic, with release planned for next year.

State-aligned hackers turn AI agents on a supply chain

Microsoft's threat intelligence team says Storm-0940, a group it assesses as aligned with the Chinese state, used ARTEX, a multi-agent AI penetration-testing framework, to automate attacks on hundreds of organizations starting as early as September. Microsoft ties the group to a compromise of ConnectWise's multi-tenant cloud backup application.

After taking control of ConnectWise application identities around October 15, the attackers' agents used those app credentials to try signing in to more than 2,250 downstream customer tenants. Earlier, between October 10 and 14, a password-spray campaign produced more than 151,000 failed logons, though Microsoft saw no successful sign-ins from it. According to Microsoft, some intrusions reached privileged cloud access in under three hours, and credentials the attackers added to applications can survive as a backdoor even after the original app is disabled. The company says its own products were not compromised. It advises customers to keep ConnectWise connections disabled, preserve logs before deleting anything, and audit app credential and role changes going back to at least October 15. The Hacker News reports that security firms Zscaler and Wiz have published their own analyses of the activity.

Anthropic puts formal proofs to work at scale

Anthropic's research team reported that Claude, working with the verification startup Theorem, produced machine-checked proofs that the compiled code behind 239 OSS-Fuzz test entry points, spread across 123 open-source libraries, does not crash on any input under 64 MB. Along the way the proof attempts turned up hundreds of bugs, including an out-of-bounds write in a decompressor. Claude patched them and then proved the fixed code.

Anthropic is careful about what this does not mean. The proofs assume a single-threaded static binary running on RISC-V under Linux, which is not how most people run this software. The authors stress that formal verification is not the same as being bug-free. The work also still takes hundreds of millions of tokens, so cost remains a hurdle.

Mega-rounds for OpenAI and DeepSeek

CNBC reports that OpenAI is negotiating a round of up to $30 billion at a $1.25 trillion pre-money valuation, led by Coatue and investors in the UAE, according to TechCrunch's summary. That would be the highest valuation ever for a private company, though it falls below the $1.4 trillion Bloomberg reported was under discussion a month earlier. OpenAI declined to comment.

Separately, Bloomberg reports that DeepSeek has raised about $15 billion, roughly 100 billion yuan, with Tencent and battery maker CATL among the possible backers. The reported raise came two days after the U.S. ordered federal civilian agencies to remove DeepSeek technology from their systems over security concerns.

Why it matters

Thursday's news shows AI agents moving out of demos and into high-stakes work, both helpful and hostile. The same multi-agent approach that improved diagnosis on hard cases also let attackers probe thousands of corporate tenants at machine speed. Meanwhile, money continues to flow to the companies building the underlying models.