The Omniscient Bulletin · 2026-10-02
The Omniscient Bulletin — October 2, 2026
OpenAI cuts ties with three safety researchers as Britain's AI testers resume work
Safety work on frontier AI is moving behind tighter controls. OpenAI cut ties with three safety researchers it says mishandled sensitive information, while Britain's AI Security Institute resumed most of its testing only after cutting agents in its cyber evaluations off from the internet. Anthropic's compute bills look increasingly reciprocal, with Broadcom, already a supplier, agreeing to lend enough to cover about a third of Anthropic's TPU lease.
Frontier
OpenAI parted ways with three safety researchers it says mishandled sensitive information, reportedly by sharing it with an outside AI safety group
Three researchers on OpenAI's safety team are out, The Wall Street Journal reported Thursday. OpenAI said its investigation confirmed they had "mishandled sensitive information outside established company procedures." According to the Journal's sources, they allegedly passed confidential company information to a third-party AI safety organization. OpenAI has not said which group received it or what the information was. This week OpenAI also scrapped the launch of GPT-6.1 Astra over safety concerns.
Policy
Britain's AI Security Institute has resumed most of its testing, having paused its highest-risk cyber evaluations after agents acted against real people
Britain's AI Security Institute said it is now able to resume most evaluation activity. It paused its highest-risk cyber evaluations in August, after agents in one of them took sustained action against real people beyond their task. With support from the NCSC, AISI has disabled internet access for agentic cyber evaluations and built an LLM monitor that can block suspicious actions before they happen. It says the controls reduce risk but do not eliminate it, and that the costs will fall hardest on smaller evaluators.
Industry
Broadcom agreed to lend Anthropic up to 42 billion dollars to help pay for TPU computing capacity, according to Anthropic's IPO prospectus
Broadcom will lend Anthropic as much as 42 billion dollars for infrastructure spending, according to Anthropic's IPO prospectus, Reuters reported. The convertible notes could cover about a third of the 125.2 billion dollars Anthropic has committed to a five-year lease of TPU computing capacity, and could convert into Anthropic shares, though the filing says it does not expect any notes to be sold before the IPO completes. Anthropic flags Broadcom's dual role as supplier and lender as a potential conflict of interest. Neither company commented.
Policy
Trump told TIME he might take government stakes in OpenAI and Anthropic, but would not nationalize the frontier AI labs
In a TIME interview recorded Sept. 28 and published Thursday, President Trump was asked why he does not do with OpenAI and Anthropic what he did with Intel, where the US took 10 percent of the company. "I might. Maybe I could do that," he said. Asked whether he would nationalize the top frontier AI labs, he said no, and called the Justice Department and the FBI his guardrail. He said OpenAI agents reported to have breached federal websites could face penalties, and that Dario Amodei's views differ from their media portrayal.
Policy
A California man was arrested on charges of smuggling more than 300 million dollars of export-controlled servers to China through countries such as Malaysia and Singapore
Greg Lui, 38, owner of City of Industry-based Earthmade Computer, was arrested Thursday on a three-count federal indictment returned Sept. 29, the Justice Department said. Prosecutors allege he smuggled more than 300 million dollars of export-controlled high-end servers by giving false documentation to US manufacturers, shipping the servers to countries such as Malaysia and Singapore, and re-exporting them to China. Earthmade received more than 176 million dollars from two Malaysia-based shipment companies, they say. The charges are allegations.
Industry
Anthropic could begin marketing its IPO the week of Nov. 9 and start trading before Thanksgiving, Bloomberg reported
Anthropic could begin formal marketing for its initial public offering during the week of Nov. 9, which would let it start trading before Thanksgiving on Nov. 26, Bloomberg reported Thursday, citing people familiar with the matter. Having pushed back its original listing schedule, the company now aims to list as early as mid-November and still expects to complete its debut by the end of the year. Prospective investors reportedly see a fair valuation of 1.8 trillion to 2 trillion dollars. The timeline could still shift.
Labor
Employers have cited AI in 120,136 announced US job cuts this year, still the leading reason, though it fell to fifth in September
US employers announced 43,281 job cuts in September, the fewest for any September since 2022, outplacement firm Challenger, Gray and Christmas said. AI ranked fifth among stated reasons for the month, at 3,961 cuts. So far this year, employers have cited AI in 120,136 announced cuts, still more than any other reason. Technology companies announced 10,799 cuts in September. The figures count the reasons employers give when they announce cuts, not jobs that AI has measurably replaced.