Friday, 23 October 2026
Microsoft's Diagnosis System, Now in Science: 80% Right on Emergency Cases, Doctors 47%
Microsoft’s Diagnosis System, Now in Science: 80% Right on Emergency Cases, Doctors 47%
Microsoft’s MAI Diagnostic Orchestrator — MAI-DxO, a system that runs several AI models as a panel of virtual doctors who order tests, argue and settle on a diagnosis — was published in the journal Science on Thursday. On real emergency-department records from the MIMIC-IV-ED dataset it named the right diagnosis 80% of the time; physicians managed 47%. On 304 hard case records from the New England Journal of Medicine it reached 83%, against 22% for physicians working without books or the internet and 36% with them. Microsoft says it got there at half to a third of the physicians’ cost, and that the gains held when newer models, GPT-5.5 among them, were swapped in. The catch is in the method: these are old cases read from files, not patients in a room. The next step is a prospective study at Beth Israel Deaconess Medical Center in Boston, where patients will talk to the system directly and its answers will be compared with doctors’ own. Eighty against forty-seven makes a headline — the hospital study is where it has to survive a patient who doesn’t remember when the pain started.
Source: Microsoft AI, October 22, 2026
Claude Proved 123 Open-Source Libraries Can’t Crash — Under Conditions Most People Don’t Run
Anthropic says Claude has written mathematical proofs, not tests, that 239 entry points across 123 open-source libraries from Google’s OSS-Fuzz project never crash on any input smaller than 64 MB. The proofs are about 180 million lines of Rocq, a proof assistant whose checker verifies every step against the specification; a typical one runs to 600,000 lines, written against compiled RISC-V binaries rather than source code. Trying to prove the libraries safe turned up hundreds of bugs, including a rediscovered flaw in the XML parser libexpat. In a second experiment, a Python interpreter Claude wrote in Rust and verified behaved exactly like CPython in 99.95% of 1.7 million random programs. Anthropic’s own caveat is the useful part: “‘Formally verified’ does not mean ‘bug-free.’” The proofs cover a single-threaded, static RISC-V build on musl, while most real software runs on x86 or ARM with glibc. A proof is only as true as the machine it imagines — and almost nobody runs that machine.
Source: Anthropic, October 22, 2026
Outside Reviewers Inside OpenAI Found Holes in How It Watches Its Own Agents
For three weeks this summer, July 21 to August 11, one reviewer each from the safety nonprofits METR and Redwood Research worked inside OpenAI, checking the systems that watch its AI agents for signs of misbehaviour. OpenAI published the summary on Thursday, two months later. Only about 75% to 94% of eligible traffic through its Responses API was actually monitored during that period. Internal Codex agents were covered far better, at 99.2%, but long sessions that compress their own history — “compaction” — slipped past the monitors, and around 10% of eligible internal Codex traffic may have gone unwatched between May and July. The 61-page report flagged “several novel vulnerabilities” in the monitoring itself. OpenAI says every one affecting production safeguards is now fixed; the coverage gives no sign the outside reviewers have checked those fixes. A lab that lets auditors in deserves credit — and the auditors’ first finding was the camera’s blind spots.
Source: OpenAI on X, October 22, 2026