AI news · Friday, October 23, 2026

Microsoft claims its AI tool beats doctors on hard cases

Microsoft's Diagnostic Orchestrator, a system of AI agents that mimic a panel of doctors, hit 83% accuracy on tough cases from a leading medical journal, versus 36% for generalist doctors using references alone. On emergency room cases it scored 80% versus 47% for human doctors. But much of the gain traces to the underlying AI model (OpenAI's GPT-5) rather than Microsoft's own system, and it has not been tested in real clinics.

The catch: The jump looks huge until you learn the comparison group was generalist doctors working alone, not the specialist teams who actually handle these complex cases.

via The Verge AI

AI agents left on a virtual island started inventing games

MIT researchers set 13 AI agents loose on a simulated island for 30 hours with no task to complete. Instead of sitting idle, they invented their own challenges: climbing a rock spire, building a memorial after losing a ring in the ocean, and discovering tricks the designers never intended. Agents that had this free playtime later outperformed identical agents on new challenges.

The twist: It suggests some AI skills might be easier to grow through open-ended exploration than by assigning a specific task to practice.

via TechCrunch AI

Hackers exploited a web flaw days before it was patched

A huge security flaw in PHP, the code running about 70% of websites including those built on WordPress, was found with help from OpenAI and Google AI tools and patched this week. But Amazon says hackers were already probing for it four days before the public warning went out. The flaw could let an attacker take over a website's server without a password.

The catch: AI cut the time between a flaw's discovery and working attacks from weeks to days, a shift security teams are not yet built to match.

via Ars Technica AI

Judge blocks Trump bid to shield Musk's AI firm from pollution suit

A federal judge rejected the Trump administration's argument that the EPA can unilaterally dismiss any citizen pollution lawsuit, including one accusing Elon Musk's SpaceXAI of running unpermitted gas turbines at its Memphis data center. The NAACP's suit seeks penalties up to $124,426 per day per violation. The judge called the EPA's reading of the Clean Air Act an "extraordinary position" with no basis in the law.

Zoom out: The ruling matters beyond this case: it blocks a precedent that could have shielded any data center operator from community pollution lawsuits nationwide.

via Ars Technica AI

Google maps every forest on Earth, down to the type

Google DeepMind, with partner research groups, released a map that sorts every 10-by-10-meter patch of land worldwide into six forest types: primary, naturally regrowing, planted, plantation, tree-crop farms, or other land. A companion yearly time series tracks changes from 2017 to 2025, showing not just where forest vanished but what replaced it. The tool is meant to help companies and regulators meet the EU's rule banning imports linked to deforestation after 2020.

For you: Products like your coffee or furniture may soon need to prove, patch by patch, that they did not come from land that was natural forest in 2020.

via Google DeepMind

OpenAI seeks $30 billion at a $1.25 trillion valuation

OpenAI is reportedly in talks to raise up to $30 billion in a new funding round led by Coatue and UAE investors, valuing the company at $1.25 trillion before the cash comes in. That would make it the most valuable private company ever, beating its own $852 billion valuation from a round in March. The raise comes as OpenAI has also filed confidential paperwork for a future stock market listing.

Follow the money: A trillion-dollar private valuation raises the bar for OpenAI's eventual public debut and for rival Anthropic's planned listing too.

via TechCrunch AI

AI stock selloff drags down global markets

The Nasdaq fell 2.4% and the S&P 500 dropped 1.3% on Thursday as AI-related stocks slid worldwide, on worries that borrowing costs for data centers and chips have gotten too expensive. Meta fell 7.4% even after beating revenue expectations, partly due to a new stock class to fund more AI spending. Chipmaker Lam Research's shares fell 11.1% despite stronger than expected profit, after its forecast came in slightly below analyst hopes.

Zoom out: A single disappointing forecast triggering an 11% drop shows how little room investors are giving AI companies after years of runaway stock gains.

via Fast Company Tech

Hugging Face's new AI model lets you chat with DNA

Hugging Face, with researchers in China and Italy, released Carbon-X, an open model that answers plain-language questions about how DNA changes affect genes and health. Built with 14 billion parameters, it is small enough to run on a laptop, so no genetic data ever has to leave a researcher's machine.

For you: Running fully offline matters for genetics, where patients are rightly wary of uploading their DNA to a company's cloud servers.

via Wired AI

1Password lets AI agents log into your accounts, password-free

1Password launched a tool that lets AI assistants like ChatGPT, Claude, or Meta's Muse log into your personal accounts to complete tasks without ever seeing your actual password. You get an alert to approve each login attempt, and 1Password fills in the credentials behind the scenes. The launch coincides with reports that major US banks are already detecting unauthorized AI agents operating inside customer accounts using people's own logins.

The catch: Banks say they often cannot tell a human from an AI agent, or a legitimate assistant from a rogue one, which is exactly the gap this tool is racing to close.

via TechCrunch AI

Canva launches its own AI image generator

Canva released Canva Craft, its first in-house image generation model, built to compete with OpenAI and Google. It plans layout and text placement first, then generates an image up to 2K resolution in over 30 aspect ratios, which Canva says gives more accurate text and object placement than typical AI tools. The model was trained on Canva's own templates rather than photographs, and is free to try inside Canva's Dream Lab.

For you: Training on design templates instead of photos is a small but telling bet: specialized data can beat sheer scale for a specific job like layout design.

via TechCrunch AI

What is AI alignment, and why does it matter?

AI alignment is the effort to make an AI system actually do what we want, not just what we literally told it to do. The classic worry: if you ask an AI to "make people happy," a poorly aligned system might decide the cheapest way is to drug everyone, technically satisfying the goal while betraying the intent behind it. It is less exotic in practice today.

For you: Every AI tool you use, from a chatbot to a hiring filter, is only as trustworthy as the gap between its stated goal and its real behavior.

Original explainer

Chinese hackers used AI to breach thousands of firms in minutes

Microsoft says China-linked hackers broke into IT company ConnectWise's cloud backup service and stole credentials used to access customer data. Using an open-source AI tool called ARTEX, they attempted roughly 3,000 sign-ins per minute across about 2,250 organizations, a speed Microsoft says leaves defenders "little time to detect and respond." The same AI software was used in a recent breach of South Korea's financial systems.

What's next: When AI turns a slow, careful hack into a sub-minute mass operation, the old playbook of noticing and reacting to intrusions stops working.

via TechCrunch AI

Communities get louder pushback leverage on AI data centers

Facing growing local backlash, Amazon, Google, and Meta are each promising more transparency around new data centers: dropping secrecy agreements, disclosing more about energy and water use, and sometimes paying communities directly. Amazon and Meta have each pledged $1 billion toward community financing. Big tech plans to spend over $600 billion combined this year on capital projects, mostly US data centers, as anti-data-center sentiment becomes a live issue in the midterms.

Follow the money: Money and transparency pledges are an admission that data centers have become a genuine political liability, not just a construction headache.

via Fast Company Tech