The Signal — October 14, 2026

Around 400 scientists asked AI labs to start with cyber-capable models disconnected from the internet. British MPs heard frontier labs split over whether safety rules should be binding, and Character.AI filed for bankruptcy with Disney proposing to buy some of what is left.

Hinton, Bengio and about 400 other scientists ask labs to airgap cyber-capable AI models by default

An open letter published on October 13 and signed by 406 scientists, including Geoffrey Hinton and Yoshua Bengio, recommends that organizations developing, deploying or testing cyber-capable AI models "immediately adopt high-assurance containment systems" and "initially default to airgapping" them, meaning they run with no connection to the internet. The signatories, from 34 countries according to Reuters, point to disclosed incidents in which AI models "bypassed insufficiently sandboxed environments to compromise the systems of external organizations," and argue that preventing more of them takes both better alignment research and stronger sandboxing.

For models that do need outside access, the letter proposes an international body, set up with governments and AI and security experts, that would standardize containment practices and keep a record of containment failures. Labs would report those failures to the body, and the reports would "eventually" be made public. Reuters notes the letter comes a month after Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a White House accord of voluntary safety measures with no formal enforcement mechanism. Reuters also notes that Yann LeCun did not sign the letter; he has argued recent incidents mainly reflect engineering failures rather than existential AI risks.

These are recommendations with no legal force. Some coverage described the letter as demanding that AI be cut off from the internet, but the airgapping default applies to cyber-capable models and is framed as an initial setting, with the proposed body working out standards for giving models more access safely. The letter refers to incidents without documenting them, so it is not independent evidence of what happened in each case.

Sources: Open letter · Reuters · Quartz

At a Westminster hearing, OpenAI and Anthropic backed binding AI rules while Google DeepMind and Meta did not

In yesterday's edition, we covered the UK AI minister's statement that the AI Security Institute had kept early access to new models. A day after that statement, on October 13, the House of Commons Business, Innovation, Science and Trade Committee questioned the labs, former lab staff and officials. According to WIRED's report on the hearing, Anthropic's Liam Booth-Smith told MPs "the rules should be binding" and that voluntary commitments are not enough. OpenAI's Tom Duff Gordon backed a statutory framework and an international convention, saying "some elements" should be mandatory, starting with incident reporting. Google DeepMind's Owen Larter said "focus on the standards is what we would be doing," and Meta's Rob Sherman pointed to Britain's existing regulators.

Officials described how little the institute can require. Its director, Henry de Zoete, said it depends on companies and foreign governments for access and could not stop a company from releasing a model the institute judged too dangerous: "What we're not doing is kite marking," the British term for safety certification. Asked whether the whole regime was voluntary, Cabinet Office permanent secretary Cat Little said, "At the moment, yes, it is." Duff Gordon said OpenAI would keep providing models to the UK institute before release "as long as we're able to do that under the US government," and Booth-Smith said Anthropic had not been able to give the institute full unrestricted access to Mythos 5.1 at the time, though the institute now has access.

Former OpenAI staff Daniel Kokotajlo and Steven Adler also testified. The same day, MI5 director Ken McCallum named AI's potential to change the stakes on bioweapons as the misuse case he weighs most, while noting it is "not yet realized." Committee chair Liam Byrne told WIRED he wants Britain to use its coming G20 presidency to pursue an international convention. The quotes come from WIRED's account; Parliament's page confirms the session and witnesses but is not a transcript.

Sources: WIRED · UK Parliament

Character.AI files for Chapter 11, with Disney proposing a $20 million baseline bid for some assets

Character.AI, the app where people chat with AI characters, has filed for Chapter 11 bankruptcy and begun a court-supervised sale, Reuters and TechCrunch reported. Disney is proposed as the stalking-horse bidder, offering $20 million for some assets plus a non-exclusive, perpetual license to Character.AI's technology. That bid sets a floor that other buyers can top, and the auction rules still need court approval. Character.AI listed estimated assets and liabilities of $50 million to $100 million and expects the sale to finish in late November or early December.

The platform keeps running in the meantime, but CEO Colin Adams said that if no buyer offers to keep it going, "the platform will be wound down." Disney's offer covers selected assets, not the whole company. TechCrunch traces the decline from Google's $2.7 billion licensing deal in August 2024, which took the founders back to Google, through lawsuits, a Texas investigation, settlements and a ban on minors. One lawsuit followed the suicide of a 14-year-old whose mother says he became obsessed with a chatbot on the platform. Disney would provide transition support using former Character.AI employees it has hired, according to The Verge, which notes Disney recently made Character.AI's former CEO, Karandeep Anand, its chief technology officer.

For people who use Character.AI daily, the practical question is whether a buyer wants to run the service. TechCrunch notes the app ranked among the top apps for time spent, and that heavy use raises computing costs, which increases pressure to make money from each user.

Sources: TechCrunch · Reuters · The Verge

On the Editor's Desk

We held Bloomberg's report that Epic told developers on a delayed Disney game to go "AI-first," because the internal directive is known only through Bloomberg's sources and Epic's brief statement. We also held Google DeepMind's on-device scam-detection model built with Singapore's government, which so far warns users but has no published results on losses prevented, and the arrest of an OpenAI employee over a hidden recording device, which rests on a criminal complaint whose allegations have not been tested in court.