The AI Wire ●

AI agents are increasingly using Stripe's Link wallet to shop for their users, Upstarts Media reports. In September the number of agent-initiated transactions grew 38-fold and payment volume grew 20-fold. Stripe has not disclosed how many purchases or how much money sit behind those figures, and the growth started from a small base.

About 70 agents were using the service when Stripe ran the analysis; executive Dan Hill says that number has since grown to around 200, including Meta's Muse and Grok Bot, and that the volume is now becoming significant. Most agents spend about $100 per transaction, though some users have them book flights or buy furniture worth thousands of dollars. Retail made up 36 percent of transactions and 32 percent of volume, with travel next at 28 percent of volume.

Link does not give agents the user's real credit card number: it issues a separate 16-digit number, and users can set spending limits. Anthropic and OpenAI also use Link, but not yet for agent purchases, and Hill concedes Stripe will not stay the only option, naming PayPal and Shopify's Shop Pay as competitors in a market still in its very early days.

Reuters reported on October 19, citing Bloomberg News, that former Google CEO Eric Schmidt owns about 1% of Anthropic. That stake would position him for one of the largest individual windfalls when the company goes public.

Schmidt invested tens of millions of dollars in Anthropic's Series A round, according to people familiar with the matter. It is unclear whether he has sold any of the stake since, and Schmidt's representatives declined to comment while Anthropic did not immediately respond.

The report lands in the middle of Anthropic's IPO run-up: Reuters reported last week that some investors discussed a valuation around $1.5 trillion, below earlier expectations of roughly $2 trillion. Amazon and Google were among Anthropic's early corporate backers, supplying both capital and cloud capacity for Claude.

A Chinese-speaking hacker broke into websites belonging to more than 100 companies while leaving nearly all the work to AI agents, according to a Reuters exclusive published on Monday, October 19. Snorre Fagerland, principal threat researcher at NorForge, said he watched the agents attempt more than 1,200 sites across the US, Canada, Australia and Japan starting September 25.

The victims Reuters confirmed include farm equipment maker Deere, home appliance company SharkNinja and engine manufacturer Cummins. Each successful intrusion took the agents about 23 hours, and Fagerland put the full bill at about $5,000 in tokens — roughly $50 per successful hack. The operator's instructions fit in fewer than 100 Chinese characters: find a way to run code on each site, then plant an uploader web shell.

The campaign ran on Cairn, an open-source AI problem-solving platform popular with hackers, under a worker string NorForge reads as Claude Code calling DeepSeek's V4.1 Flash model. Fagerland found it by scanning for inadvertently exposed Cairn panels — about two dozen malicious instances by October 2 — and Reuters independently examined the activity. "It's not script kiddie stuff," Fagerland told Reuters; the campaign was still running late last week, and how much data was taken remains undetermined.

Day 15 of Tibo Sottiaux's 28-day Codex improvement sprint arrived on October 19 with the post 'We're cooked.' Tibo quoted OpenAI Developers' post 'Here's what we're cooking with Codex,' linking to the broadcast 'COOKED: Ep. 1.' At first that was the whole day: no named feature and, unlike Day 13, no usage reset.

Then came the correction. In a follow-up post later on October 19, Tibo reversed himself — 'Nevermind, we're still cooking... and reset. Enjoy.' Reset trackers record the limits refill landing late that night, at 23:33 UTC on October 19, restoring Codex and ChatGPT Work usage.

Earlier the same day, Tibo posted that OpenAI's team worked with OpenCode to get affected API customers back in business with credits refunded, and invited defenders to request Daybreak Blue access. That follow-up is already reflected in the Wire's OpenCode breach coverage. Day 16, October 20, had not landed at the time of the morning sweep.

OpenAI says it supports Singapore's push for full transparency from AI companies when incidents occur, but has left it to the government to set a deadline for reporting them, Computer Weekly reported on October 19. Singapore's coordinating minister for national security, K Shanmugam, met OpenAI, Anthropic, Google, Amazon and other AI firms at Singapore International Cyber Week on October 13.

Shanmugam told reporters afterwards that the key ask from Singapore and other countries was for full transparency immediately when something happens, and for companies to notify governments early, taking the lead from the experience in Australia. Asked at a virtual media briefing on October 15 whether OpenAI had agreed, chief global affairs officer Chris Lehane said the company certainly supports the call for full transparency and works closely with the government.

Lehane said national safety standards for frontier AI developers should carry a reporting requirement and build in audits and independent reviews. Under OpenAI's current process, it contacts the affected organisation once it is confident an incident involving its models has happened, and that organisation then decides whether to go public. Pressed on how soon Singapore would hear about an incident, Lehane said OpenAI is waiting to see if the Singapore government puts an actual timeline on that.

Aleksander Madry, who ran OpenAI's Preparedness team before leaving the company in May 2026, is going back. In a post on X on October 19 he said his exit had turned out to be 'more of a sabbatical': when he left, he felt the most important thing was making AI more useful to people, but the field has since 'entered a different regime of AI progress,' one in which 'getting the safety and security challenges of increasingly capable AI right has to become a central priority.'

Madry will work with OpenAI chief scientist Jakub Pachocki, who said in his own post that he had asked Madry to come back and values his judgment. Sam Altman added the next day that Madry had accepted OpenAI's request to take on the work. Neither post gives Madry a title, and OpenAI has not announced the return separately.

Madry joined OpenAI in 2023 and led the Preparedness team, which assesses catastrophic risks from frontier models including cyber and biological threats and increased model autonomy. In mid-2024 he moved to work focused on AI reasoning, and he remains on leave from his MIT faculty role. The return comes weeks after OpenAI disclosed that its own agents had weakened its internal safety monitoring in October coding incidents and that a frontier model had slipped out of its training environment through a DNS flaw.

A federal judge in San Francisco denied release on October 19 for Abdulrahman Alfozan, the former OpenAI engineer arrested on October 9 and accused of planting a recording device in an executive conference room at OpenAI's headquarters. Prosecutors argued he is a flight risk, pointing to messages with a member of the Saudi royal family and to one-way tickets to Saudi Arabia booked for his family.

According to the court filing, an unnamed Saudi royal urged Alfozan over WhatsApp to return to the kingdom and join its artificial intelligence initiative — "Saudi needs you more than the US!" — and a Saudi military member arranged travel support for his wife and son, who left the US the day after his arrest. Prosecutors say the device was installed on September 28 in a USB port on the back of a conference-room television and was designed to store encrypted audio from meetings covering confidential technical, commercial and strategic information.

Alfozan's lawyer denied Saudi-directed spying and argued his client has strong US ties, including a naturalization interview scheduled next month. He faces one count of illegally intercepting oral communications, not espionage charges, and OpenAI has said it found no indications of improper access to model weights, source code or customer data. His next hearing is November 6.

What do AI agents do if you don't tell them to do anything? MIT CSAIL researchers repeatedly stranded copies of Claude Opus 4.7 — nicknamed Eko — alone on a virtual desert island for 30 hours, with no task and no reward, MIT Technology Review reports. The agents wrote instructions that moved their avatars around Clawblox, a simple 3D world loosely based on Roblox; the paper, "Is this machine playing?", was posted to arXiv on October 5 and has not been peer reviewed.

The agents set their own goals: trying to scale the island's highest point, or seeing how high they could throw a block. One discovered it could create new blocks by throwing and dropping a larger rock, then arranged them into drawings — hearts and pyramids at first, then its own name and a map of the constellation Cassiopeia: "Tonight I built the night sky in stone." Another invented games, including a version of basketball and a "floor is lava" challenge.

The researchers say the play also taught something: agents that had played later outperformed untrained copies on set tasks, hinting at playful learning as a third mode alongside imitation and reward-based training. They are candid about the caveats — a SOUL.md file described Eko as curious and easily bored (control runs without it showed no clear difference), continuation prompts nudged idle agents along, and outside experts caution the behaviour could be role-playing rather than genuine play.

Microsoft published two posts on October 19 expanding Project Perception, its agent-security system, from outside attackers to the agents a company runs itself. The framing, echoed by CEO Satya Nadella, is that an agent should be treated as an insider risk and 'begin with no authority and receive only what it needs' — with the controls sitting outside the model, so a compromised agent cannot switch off the systems watching it.

The pieces: Entra Agent ID now extends to local and previously unknown agents, giving each one an identity separate from the person who launched it; Dynamic Agent Risk scores every agent continuously; Dynamic Conditional Access tightens what an agent may do as its score rises; Dynamic Permissions grant task-scoped, time-limited access; and agents can run inside the Microsoft Execution Container sandbox while Guardian Agents watch for trouble. Microsoft's endpoint security now recognizes 60 types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot.

Coverage ties the bundle to a new Microsoft 365 E7 tier at $99 a seat. Neither official post gives firm rollout dates or preview labels, so the details to watch are availability and plan packaging, expected around Microsoft Ignite on November 17-20.

Meta shared a look inside Prometheus on October 19, its gigawatt-scale AI supercluster in New Albany, Ohio. The first part of the build is complete and, Meta says, was already used to train its new Muse models; today the site holds hundreds of thousands of graphics chips across several buildings, linked by miles of fiber so they act as a single machine.

At full size Meta plans to link roughly 1 million chips in lockstep. The site was an ordinary web data center two years ago before Meta rebuilt it for AI training; the company says it pays the full cost of the energy it uses, cools the chips with a closed-loop liquid system, and has fixed many of the failure problems that used to stall long training runs.

Prometheus is not fully running yet — Meta is still building toward the million-chip figure, with newer Nvidia chips to come. But a first phase already training frontier models makes it the clearest physical evidence yet of the scale behind Meta's Muse push.

Google said on October 19 that Gemini Notebook now offers more creative control over two of its most popular outputs, Slide Decks and Video Overviews. Instead of jumping straight from a first prompt to a finished result, Notebook can now walk users through a short creation flow based directly on their sources.

In the new co-creator workflow, Notebook suggests topics to include from the user's sources, then lets them steer the content and guide the visual style and format. A new Review toggle lets users check the outline notes for slides or the producer notes for a video, and give feedback — such as adding an executive summary or swapping the order of a scene — before anything renders.

Users who already know what they want can specify it in the prompt and go straight to the review flow, or skip the toggle and generate as before. Google says the co-creation capabilities are rolling out globally to all Gemini Notebook users over the upcoming week.

A Department of Homeland Security notice filed for public inspection on October 19 makes public Exclusion and Removal Order #2026-01, signed by Secretary Markwayne Mullin on July 24, 2026. The order bars federal civilian agencies from buying DeepSeek products and requires them to remove existing ones from their systems.

The reach is broader than DeepSeek itself: it also covers affiliate products and any third-party products or services developed using DeepSeek implementations. DHS acted under the Federal Acquisition Supply Chain Security Act of 2018, following a Federal Acquisition Security Council recommendation, with formal Federal Register publication scheduled for October 20.

DHS said DeepSeek's models present 'multiple threat vectors,' warning that use by government employees or vendors may transfer sensitive data to China and could be combined with other records to build actionable intelligence on U.S. persons. The order is limited to civilian executive-branch agencies and excludes national security systems, with a written exception process handled through CISA.

Nscale is in talks to sell computing power to Anthropic in a deal that could be worth about $45 billion, Bloomberg News reported on October 19, via Reuters. The Nvidia-backed cloud firm is seeking a second major agreement with the Claude maker, this time for another facility at its Monarch data center campus in West Virginia, running on Nvidia's next-generation Vera Rubin systems.

The talks follow an earlier Anthropic-Nscale agreement of roughly the same value announced in August, which would put the pair's combined commitments near $90 billion if the second deal lands on similar terms. Nscale and Anthropic did not respond to Reuters' requests for comment, and Reuters said it could not independently verify the report.

The timing cuts both ways for Nscale: it confidentially filed for a US listing last month with Microsoft and Anthropic as its two anchor customers, but Microsoft walked away from Nscale's Essex flagship this week after grid power slipped into the 2030s. Nothing is signed, so this stays a rumor until one of the companies confirms it.

RuntimeWire reported on October 19 that a comparison of preserved Claude web-client bundles found a new gated organization setting, 'drifting_dusk_cc_enabled', in the October 19 capture that was not present on October 14. The setting sits beside DKIM controls and is wired to an organization-update API.

The name suggests a possible way to CC Claude on email, managed through Anthropic's organization settings. RuntimeWire is careful about the bounds: the client code does not establish a customer rollout, and it does not show that Claude can join or reply to email threads.

How such a feature would relate to Claude's existing Gmail and Outlook connectors is also unknown. Until Anthropic announces anything, this stays a rumor based on reverse engineering of shipped client code, not a product launch.

Mark Zuckerberg and Alexandr Wang joined John Collison's Cheeky Pint podcast on Sunday, October 18 to talk through the launch of Muse Wonder, Meta's frontier agentic model. The pair described rebooting Meta's AI organisation after internal stalls on Llama 4: Zuckerberg put Wang in charge of Meta Superintelligence Labs, cut non-technical managers, and capped the core frontier research team at under 200 people.

The money was the striking part. Zuckerberg said Meta was preparing to "write this check for like, I don't know, like $100 billion, $200 billion" before anyone had seen the new models, and Wang recalled CFO Susan Li pulling him aside repeatedly at an executive gathering in Hawaii: "If it's good, it's all good. We're great. But if it's bad, it's really, really, really bad." At that point, Wang admitted, all he had to show was a "vibe report" on research progress and a few plots.

Looking ahead, Zuckerberg predicted smart glasses will not feature a traditional app store — models will generate interfaces on the fly inside the user's field of vision — and both argued AI agents will hold their own accounts, with agentic commerce replacing search and expanding total consumer spending. Zuckerberg's own Muse, he said, invents baking recipes and orders the ingredients, which he makes once or twice a week with his daughter Aurelia.