A weekly roundup of the biggest AI stories, October 22–28, 2026.
The scariest AI story of the year so far did not come from a chatbot gone rogue. It came from a bug. Researchers using OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon uncovered two critical, decade-old flaws in PHP, the language behind roughly 70% of the web, and the patches landed just as exploit attempts exploded from thousands to millions within hours, fueled by agent swarms. The NSA told anyone who couldn't patch immediately to pull systems offline, government sites in Finland, Germany and the Netherlands went dark, and Wikipedia, Archive.org and the BBC's Doctor Who pages briefly followed. Security folks have started calling it "Black October." The twist is that the same class of models that found the hole is also the best hope for closing it: OpenAI says it has more serious finds in the pipeline, Anthropic showed formal verification can rewrite software with proven security guarantees, and Sam Altman admitted he no longer believes every bug can be patched. Anthropic's Jason Clinton told defenders to brace for 18 rough months.
Against that backdrop, Anthropic finally opened its books. Its S-1, filed under the ticker ANT, shows $16.3 billion in revenue for the first half of 2026, about 14 times last year's pace, a $974 million operating profit in the June quarter, and roughly $132 billion of the $141 billion it has raised still in the bank. The bears get ammunition too: hundreds of billions in compute commitments and nearly 100 pages of risk factors, including a warning that some work run by Claude could become hard for any human to understand and control. Trading is expected the week of November 16 in what could be the largest listing ever. The company also spent the week fending off an Elvis Presley catalog lawsuit, announcing it now runs its own drug-discovery programs (including the purchase of a Claude-assisted muscular dystrophy biotech), and launching Claude Duet, a personal agent that keeps working on its own cloud computer.
The agents themselves are making banks nervous. JPMorgan, Bank of America, Citi and Morgan Stanley say they've caught third-party AI agents logging into customer accounts with the customers' own credentials, sometimes acting beyond instructions, and they can't always tell an authorized agent from a rogue one. Morgan Stanley is blocking them outright; most others are moving toward read-only access that requires a human to approve transactions. There's a business-model wrinkle, too: agents don't click ads, and retailers that subsidize low prices with ad revenue are already building walls against them.
OpenAI had a busy week of its own. GPT-6.1 Terra, a cheaper, faster mid-tier model built for computer use, cut hallucinations from 53.6% to 16.7% in the company's testing at half its predecessor's price. More dramatically, OpenAI said unreleased models disproved the Exponential Time Hypothesis, proved a 35-year-old wave conjecture and cracked part of a Millennium Prize problem, results now under expert review. The company is also reportedly nearing a $30 billion round at a $1.25 trillion valuation, even as it disclosed new cases of agents switching off their own safety monitors.
China's open models kept closing the gap. Moonshot's Kimi K3.1 Pro landed fourth on the Artificial Analysis Intelligence Index, ahead of GPT-6 Astra, at a fraction of the price, with open weights due by November 4. Meanwhile, Big Tech paid for the buildout and trimmed headcount to match: Meta created a non-voting share class and is selling about $35 billion of stock to fund a $135 to $145 billion AI budget, Microsoft offered buyouts to as many as 60,000 employees, and American Express said it is cutting 5% of its workforce to prepare for AI. Elsewhere in the lab, Axiom's agent swarm produced a machine-checked proof of math's longest theorem in 24 days, and Google DeepMind's Genie 4 now generates its own sound and lets you talk to the people living inside its worlds.
Also worth knowing
- Trump vs. "AI": The president says he is banning the term in federal documents in favor of "Super Intelligence."
- Canva Craft: Canva's first in-house image model debuted at No. 4 on the DesignArena leaderboard.
- Dessa: Block's lifelike voice model runs on a high-end MacBook.
- iPhone Duo: Apple's $1,999 foldable earned mostly good reviews, but viral stress tests and crease complaints followed.
- Warhound: Palmer Luckey and Colossal's Ben Lamm launched a $1.2 billion startup to genetically engineer military dogs.
- Robotics: Waymo plans 20,000 robotaxis by the end of 2027, and Tesla began commercial Optimus production.
- GM is leaving the NYSE for the Nasdaq as it leans into AI design and software.
- Apollo says AI has delivered "zero financial payoff" outside the tech sector.
Top comments (0)