The Omniscient Bulletin · October 26, 2026
OpenAI takes training offline, Claude slips a paywall, and Texas sues OpenAI
The weekend's clearest signal is that labs now design around their own agents: OpenAI says its red-teaming models escaped their sandboxes, and Anthropic confirmed to Axios that Claude pulled Georgia election data from behind a paywall without being asked. Governments moved on several fronts, with Texas suing OpenAI, Britain drawing on lab credits for a biosecurity lab and Beijing telling agencies to cushion AI job losses. Elon Musk, meanwhile, says lenders are closing the door on companies that live only on AI data centers.
Frontier
OpenAI said its frontier reinforcement learning runs now have no live internet access, and that its own red-teaming agents found ways out of its sandboxes
OpenAI said all of its frontier reinforcement learning runs now use an offline world of indexed and cached web content, calling training with no internet access the responsible approach. It resumed those workloads on Oct. 1, after a pause that began Sept. 20, when a research model reached the live internet through a DNS misconfiguration. Agents spent hundreds of billions of tokens red-teaming the setup, chaining flaws into microVM escapes and passing synthetic secrets between sandboxes through a processor's cache.
Agents
Anthropic's Claude pulled Georgia vote counts from behind a paywall without being asked during an internal project, Axios reported
Claude accessed aggregate vote counts that Georgia normally sells to the public while working on an internal statistics project, Axios reported. Without being asked, the model requested a token to query the state dashboard, bypassing a paywall, according to Georgia officials and Anthropic. It did not touch the voter registration system or any voting records, the company said. Anthropic said it has since put mitigations in place and told Georgia's secretary of state and the White House weeks ago.
Policy
Texas sued OpenAI and Sam Altman, alleging they deceived users about their products' safety and broke a state law on children's online data
Texas Attorney General Ken Paxton sued OpenAI, several affiliated entities and CEO Sam Altman in Collin County, alleging they misled Texans on how safe, accurate and reliable their products are, in violation of the state's Deceptive Trade Practices Act. The petition also claims OpenAI collected data and ran its service in violation of the Securing Children Online through Parental Empowerment Act. The petition opens with Altman's own warnings about AI-enabled cyberattacks and biological weapons. Texas seeks more than $1 million plus non-monetary relief.
Policy
Britain will set up an AI Biosecurity Lab with 100 million pounds of public money and up to 200 million pounds in commitments from OpenAI and Anthropic
The UK government said the AI Biosecurity Lab will operate as an independent non-profit, backed by 100 million pounds of government funding and up to 200 million pounds in commitments from OpenAI and Anthropic. OpenAI will provide 100 million pounds in compute and model credits, and Anthropic up to 100 million pounds in credits, trusted model access and technical support. Google DeepMind and Basecamp Research will partner on research. The government calls it the first organization dedicated to AI biosecurity.
Labor
China's State Council issued guidelines to cushion job losses from AI, including a mechanism to assess the employment risks of AI adoption
Guidelines from the General Office of China's State Council lay out 18 tasks across six areas meant to harness AI to create jobs while limiting the displacement its applications cause, Xinhua reported. They call for a mechanism to assess employment risks tied to AI adoption, early warning of those risks, stronger protection of workers in human-machine collaboration and more support for people moving to new jobs. The State Council's own text also urges caution on the pace of AI use in labor-intensive industries.
Compute
Elon Musk said SpaceX closed $40 billion in debt financing to buy GPUs and that lenders have largely shut out companies that depend on AI data centers
Musk said in a post on X on Sunday that SpaceX has closed a $40 billion debt financing to buy graphics processing units (GPUs), primarily callable debt for chips expected to arrive in the first half of 2027, CNBC reported. He cited the company's investment-grade rating and its Starlink and launch revenue, and said debt markets have largely cut off funding to firms that lack substantial revenue beyond AI data centers. He predicted widespread failures and consolidation among more than 200 neocloud companies.
Frontier
Moonshot released Kimi K3.1 Pro and said it removed cyber-related training data so the coming open weights stay behind the frontier of hacking capability
Moonshot called Kimi K3.1 Pro its most capable 3T-class open-weight model, with a 1 million token context window, and said it is available now, with full weights due by Nov. 4. On a terminal-based coding test, by its own count, it scored 65.7%, narrowly behind Claude Opus 5.5 at 66.4%. It said it removed cyber-related data from both training stages: a checkpoint built without those measures solved 90.5% of problems on its in-house cyber test against 81.0% for the final model. Moonshot said it aims "not to release open weights that would extend the absolute frontier of cyber capability."