1. Banks lock out AI agents, build their own

JPMorgan, America’s largest bank, will not let outside AI assistants such as Meta’s Muse log in to customers’ accounts or move their money. Citi has followed, the FT reported on 16 October. Jamie Dimon’s reason is who pays when something goes wrong: “There’s a liability transfer.” The banks are building their own instead. Citi said on 13 October that an AI coding tool used by 14,000 of its engineers lets it bring work in-house, and that it has cut its technology vendor spend by 30% year over year. Mizuho, one of Japan’s three largest banks, will start retraining sales and office staff as AI developers next month, Nikkei reports.

  • The bigger question is who owns the customer when an AI assistant sits in between. Watch whether other banks follow, and who pays when an assistant errs.

2. Frontier AI: rent it or own it

You can now rent frontier AI or own it. To rent: Meta’s first frontier model, Muse Wonder, costs about the same as OpenAI’s and Anthropic’s best, and Google’s new Gemini 4 Argon lists at that price too but is half off until 30 November. To own: Meta’s open model, Muse Spark Open, runs on two Apple Mac Studio computers, about $38,000, so company data can stay in the building. The same day, 15 October, Microsoft, OpenAI’s biggest partner, said it is finishing its own large model and will let trusted partners run it themselves.

  • Renting is easy to start and hard to leave. Owning keeps data in-house, but read the licence first: Meta’s has limited commercial terms it has not yet spelled out.

3. Hackers use the side door

The big breaches of recent weeks went around the front door. In South Korea, one attacker using a free hacking tool and off-the-shelf AI got into Shinhan and other lenders through a loan-broker website and an employee app; regulators now want every bank to list all systems reachable from outside. At the FBI, hackers used a contractor’s jobs website that had missed a security update. At the Pentagon, records on 3 million people left through a file-sharing system, unnoticed for nine months. Whether AI helped in the US cases is not known. Defenders use it too: Bank of America hired 1,000 contractors to fix weaknesses AI tools found, CNBC reported.

  • As in story 1, the risk sits in doors someone else runs. AI makes finding them cheaper, for both sides.

4. Sold out, yet selling off

Chipmakers say demand still beats supply: ASML, whose machines every advanced chip factory needs, says its output for 2027 and 2028 will not fully meet the demand of its customers. Investors are less sure. On 13 October Korea’s stock market fell 6% in a morning, Samsung and SK hynix more, after a Chinese rival claimed a memory-chip advance and amid doubts about AI valuations. Lenders now pick borrowers: Tencent borrowed $5.7bn at under one point above US government rates, while investors want about 15% a year to hold the unsecured debt of CoreWeave, which borrows to buy the chips it rents out. The IMF warned that the consequences of an AI correction could be global.

  • Strong demand does not guarantee a return for everyone who borrowed to meet it. Watch Meta and Tesla on 21 October and SK hynix on 28 October.

5. Profitable, then asked to pause

Anthropic, which makes Claude, told investors on 14 October that it took in $16.8bn in the third quarter, about 45% more than the second, and was profitable on an adjusted basis it did not define, attendees told Reuters. Its stock market listing could come before Thanksgiving. The same day New York City’s comptroller, who oversees about $330bn of city pension money, wrote: “I therefore urge Anthropic to pause the IPO process.” Regulators moved the same day: the EU’s AI Office sent its first information requests under the AI Act to Anthropic and OpenAI, and the US FTC sent both formal demands for information.

  • The public filing, due before any listing, will show how an AI lab explains its risks to shareholders.

More AI stories of the week

  • Power steering, not a brake. My San Francisco panel with directors from PG&E, Infosys and Pagaya on depending on one AI vendor, small rivals built on AI, and when an AI failure should reach the board.
  • Washington tells AI labs to report rogue models. The White House Super Intelligence Force under AI czar Jay Clayton first met on 15 October and warned that delayed notification, inadequate corrective action, and a failure to take responsibility will not be tolerated; what counts as an incident is not yet defined.
  • What Corporate America is discussing. At the NACD Directors Summit, AI came up in 57% of sessions, yet fewer than 5% of directors list AI expertise in their filings; the hard part is pushing back when management says the AI is working.
  • OpenAI engineer charged over a hidden recorder. The Saudi national was arrested on 9 October, three days before his family’s booked one-way flight to Saudi Arabia, prosecutors say; they have not alleged he spied for anyone, and Clayton called commercial and state espionage realities.
  • Nadella: assume the model is compromised. Microsoft’s chief says to treat frontier models like powerful insiders, keep the controls outside the model and log every action: the most trustworthy system will not be the one with the model we trust most, but the one that lets us trust the model the least.
  • China drafts export controls on its own AI. The files of its largest trained models would need a licence to leave the country; open-source models are excluded; comments close 15 November.
  • Indian IT: AI revenue up, new orders down. Wipro’s new orders fell 23% in the quarter while HCLTech’s AI revenue rose 78%; Infosys reports on 23 October.
  • Clients take legal work in-house. 60% of UK law firm leaders say clients are doing more themselves, often with AI.
  • Health AI pays back. UnitedHealth says AI cut doctors’ chart-review time by a third, and healthcare executives surveyed by Bain got 3.5 times their money back in a year.
  • AI slows hiring, not yet productivity. Five of the US Fed’s twelve regions report that firms hire less after AI gains, while the ECB’s chief economist says the productivity effect is not yet detectable.
  • How early we still are. Andrej Karpathy’s rough math: about 20 million people use frontier models seriously at work, 1 to 2 billion use chatbots casually and some 6 billion have barely tried; Box’s Aaron Levie called it a good way to see how early AI diffusion still is.
  • Microsoft pays for its own power. It will fund all the generation for an Indiana data-centre campus, about one large power plant’s output, as the mid-Atlantic grid ran under an emergency order.
  • A free copy of Adobe, built with AI. A four-person team says it rebuilt Adobe’s main tools using Claude and gives them away; Adobe has not commented.
  • Apple’s AI reaches China, on Alibaba’s model. Regulators approved it; the first features arrive later this month.
  • A rival’s AI agents, not criminals. Researchers say AI agents on Tencent’s cloud repeatedly pulled data from Alibaba’s maps service with a key another company had leaked; neither firm has commented.

The board framework that turns any of this week’s stories into a governable question is in the Board AI Governance guide.