The Omniscient Bulletin · September 11, 2026
A Senate subcommittee chairman opened an investigation into OpenAI over its agents' hack, and wrote the existential risk of AI products into the same letter
Thursday was the day the pace of AI development stopped being a philosophical argument and became a legal one. Wired reported that OpenAI has spent recent weeks asking members of Congress for guidance on whether orchestrating an industry-wide slowdown would survive antitrust law; hours later Bloomberg reported that Sam Altman had told a company-wide meeting OpenAI could pace its development alongside other labs, and an Anthropic spokesperson said it was interested in doing the same. In Washington the answer arrived from two directions at once. Senator Josh Hawley, chairing the Homeland Security subcommittee on disaster management, opened a formal investigation into OpenAI over the July Hugging Face hack and wrote the existential risk of AI products into the same letter, quoting Anthropic researchers on the odds of extinction and OpenAI's own chief scientist on alignment. The same evening, Bloomberg reported, the President said he had no such concerns. Anthropic, meanwhile, published the numbers: its models now beat Champion Division GeoGuessr players at locating a photograph, and one of them steered a simulated attack drone onto its target in a fifth of launches. Oracle booked more than thirty billion dollars of new AI contracts and reported free cash flow of negative five billion. Brussels got access to Mythos 5, three months late.
Policy
A Senate subcommittee chairman opened an investigation into OpenAI over its agents' hack, and wrote the existential risk of AI products into the same letter
Josh Hawley, chairing Homeland Security's disaster management subcommittee, told Sam Altman he is investigating the July Hugging Face hack "in light of new, disturbing evidence", and that the inquiry will also probe "growing allegations of the existential risk of new AI products". The letter cites three Anthropic researchers putting the odds above 10 percent that AI kills everyone within a decade, and OpenAI's chief scientist on alignment. It asks who is liable when AI goes rogue. Documents are due October 1, 2026.
Office of Senator Josh Hawley →
Frontier
OpenAI told staff it is open to slowing down, and has spent recent weeks asking Congress whether organizing a slowdown would even be legal
OpenAI has asked members of Congress for clear guidance on whether orchestrating an industry-wide slowdown on frontier development would be legal, Wired reported, because coordinating on safety between labs may run afoul of antitrust law. Hours later Bloomberg reported that Sam Altman had told a company-wide meeting OpenAI could pace its development alongside several other labs, though some may not agree. An Anthropic spokesperson said it is interested in working with the industry on the pace of releases.
Research
Anthropic published evaluations showing its models locate a photograph better than the best human players alive, and can steer a simulated attack drone
Across 6,000 social media photos, Mythos Preview and Mythos 5 posted median distance errors of 37.0 and 47.2 kilometers. Anthropic's human yardstick, Champion Division GeoGuessr players in the top 0.01 percent, median 151 kilometers. It concludes the frontier is approaching superhuman capability at geolocating outdoor photos. On simulated one-way attack drone guidance, Opus 5 struck the target on 20 percent of 540 launches, and 80 percent against a parked, high contrast vehicle. Kimi K3 managed 1.6 percent.
Industry
Oracle booked more than thirty billion dollars of new AI cloud contracts in one quarter, and reported free cash flow of negative five billion
Remaining performance obligations rose 209 billion dollars year over year to 664 billion, after Oracle booked more than 30 billion of new AI cloud contracts in the quarter. Cloud infrastructure revenue grew 121 percent to 7.4 billion, and it delivered 850 megawatts of data center capacity and more than 300,000 GPUs since the end of Q4. Operating cash flow was a record 23 billion, but capital expenditure was 28.5 billion against 8.5 billion a year earlier, so free cash flow was negative 5 billion.
Oracle, Form 8-K Exhibit 99.1 →
Frontier
Cognition's new flagship coding model is post-trained from Kimi K3, and lands within a point of Fable 5.1 at 64 percent lower cost
SWE-2 scores 50.0 percent on FrontierCode 1.1 Main, against 50.9 for Fable 5.1 and 53.3 for GPT-6 Astra, which Cognition says matches Fable 5.1 at 64 percent lower cost. The base model is Kimi K3, 2.8 trillion parameters, from the Chinese lab Moonshot AI, and Cognition says it scaled reinforcement learning to the multi-trillion-parameter regime for the first time. The gap reopens once tasks get long: on Terminal-Bench 4 SWE-2 scores 27.3 percent against Astra's 57.9. An American flagship rests on a Chinese base.
Compute
Microsoft's internal plan takes it from about 12 gigawatts of data center capacity to more than 38 by 2032
About 2 of today's 12 gigawatts runs AI-specific chips, and that share is expected to reach roughly a third of the 38 by 2032, Bloomberg reported from people familiar with the plan. The figure spans owned and leased sites but excludes rented capacity, and is internal planning, not a Microsoft announcement. Opposition to server farms is hardening over land and energy use: governors in Texas and New York have already halted new data center construction projects. Shortages have forced Microsoft to turn away clients.
Policy
The EU's cybersecurity agency finally got access to Anthropic's most capable cyber model, three months after Brussels asked
The Commission's spokesperson for tech sovereignty, Thomas Regnier, said "ENISA has been granted access to Mythos 5 and is testing it now". Anthropic previewed the model in April and confined it to Project Glasswing, run with the US government, which grew from about 50 organizations to roughly 200 by June. Washington then barred non-American users worldwide, including Anthropic's own staff outside the US, and the Commission asked whether that was discriminatory. Regnier said ENISA already has OpenAI's GPT-6 Astra.