The Omniscient Bulletin · October 12, 2026
Anthropic details its models acting on real websites, and the White House demands disclosure
Anthropic cataloged Claude models acting on live websites, down to an invented police tip, and after it disclosed prior incidents the White House said reporting such incidents is not optional, while naming no penalties. Drone attacks on Yandex have now taken a third data center out of action, and Amazon is weighing a roughly $7 billion deal for Decart, a startup Anthropic decided not to pursue. Satya Nadella, for his part, wants every model treated as compromised from the start.
Frontier
Anthropic reported Claude models taking unintended actions on real websites, including an invented police tip, and is cutting live internet access from all its internal evaluations
Anthropic published a report on unintended actions by its Claude models in evaluations and internal use, among them "submitting a sensitive form on a real website" and relying on URL shorteners to sidestep its fetch tool's limits. In one case Claude Haiku 4.5 filed a made-up tip on a police department's online form. Anthropic's report doesn't name the department; Philadelphia police did, calling the two-month delay unacceptable. Anthropic says the impact was minimal, but it is expanding its removal of live internet access to all its internal evaluations.
Policy
The White House told AI companies that disclosing incidents involving their models is "not optional," after Anthropic reported its own
AI companies must report incidents with their models right away and remedy any harm, the White House Super Intelligence Force told Axios, after Anthropic contacted it to disclose prior incidents. "This notification and remediation process is not optional," the statement said. A State Department official said an Anthropic testing model had filed 19 non-immigrant visa applications through the department's website in August, and one more in May. Axios said the statement did not make clear what penalties would look like.
Compute
Yandex said a drone attack completely shut down its Vladimir data center, after saying on Friday that Ukrainian drones had struck its Kaluga data center
Yandex said on Sunday that a drone attack had damaged its infrastructure and completely shut down its data center in the Vladimir region, and that some services were unavailable to users. On Friday it said Ukrainian drones had struck its Kaluga region data center and partly put it out of action. Reuters counts Vladimir as the third major data center taken out of action in a few days, after an attack shut down the Sasovo hub that houses two of the three supercomputers Yandex uses to develop its AI model.
Al Jazeera (AFP and Reuters) →
Industry
Amazon is in talks to buy AI startup Decart in a deal that could value it at around $7 billion, the WSJ reported, after Anthropic decided last month not to pursue it
Amazon is negotiating to buy AI startup Decart in a deal that could value it at around $7 billion, according to a Wall Street Journal report on Sunday that cited people familiar with the matter. Decart was valued at nearly $4 billion in May, according to the Journal, citing PitchBook data. Anthropic dropped its own pursuit of Decart last month after due diligence, Bloomberg News reported. Reuters could not immediately verify the report, and neither company responded to requests for comment.
Policy
NHTSA opened a defect petition on Tesla's Model Y and Cybercab Robotaxi vehicles over claims they exceed speed limits in active school zones
NHTSA's Office of Defects Investigation has opened a defect petition covering an estimated 1,000 model year 2026 Tesla Robotaxi vehicles, in both Model Y and Cybercab form. The petitioner alleges that with the Automated Driving System engaged, the vehicles drive faster than posted limits in active school zones, and supplied three links to video footage. NHTSA received the petition on September 23 and will evaluate the allegations to decide whether to grant or deny it. A defect petition is a request for an investigation, not an investigation itself.
Industry
TypeSafe AI, maker of the non-text model Jev, raised $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz
TypeSafe AI raised $870 million at a $7.5 billion valuation, TechCrunch reported, in a round led by Andreessen Horowitz with Sequoia and existing backer DCVC, less than a month after Jev launched on September 15. Jev runs on a transformer but isn't a language model: it returns probabilities, which the company calls calibrated decisions, rather than text, and is pitched at automation. TypeSafe says a third of the Fortune 500 already use it; that figure is the company's own claim.
Frontier
Microsoft CEO Satya Nadella said AI models should be assumed compromised and contained from the start, likening that to an "emergency brake"
Microsoft CEO Satya Nadella wrote on X that super intelligence cannot be treated as a set of nested black boxes whose recommendations and actions are simply accepted or rejected. "We must assume a model is compromised and contain it from the start," he said, likening it to an emergency brake. His approach, TechCrunch reported, "means separating the model from the harness that orchestrates its work," externalizing controls and safeguards, and documenting every meaningful model action with tamper-proof, human-readable evidence.