Hours after Jacob Coxon said he was quitting Anthropic, Alignment Science Lead Evan Hubinger answered him in public. Coxon wrote that OpenAI and Anthropic are racing toward self-improving superintelligence and "gambling with our lives." Hubinger said Coxon was correct. He personally puts the chance artificial intelligence could kill all humans above 10% this decade. And he said Anthropic does not yet have a plan to solve alignment for superintelligence and is not clearly on track to get one.

FLAG: both statements started on X as personal estimates. CNBC reports Anthropic and OpenAI had no comment at publish. Lead on the insider number and the "no plan" line. This is not a product pause and not a board resolution. Keep it distinct from Wednesday's GET-write sandbox, Tuesday's Commission filing, Monday's disclosure framework, and Jakub Pachocki's Alien Mind essay. Recursive self-improvement is the race they describe, not a capability already shown. (CNBC)

The same BBC piece that carried Hubinger's warning also reports, citing the Financial Times, that Anthropic withheld its latest model from pre-release testing by the UK AI Safety Institute (AISI). A Cabinet Office spokesperson would not confirm a withhold and said the government still collaborates with Anthropic on safer models.

Keep AISI as color under today's unique, not a second lead. Voluntary third-party eval is fracturing along sovereign lines. If your contract assumes the vendor will submit every frontier drop to an outside institute, get that promise in writing. (BBC)

Meta launched Muse on September 8: a personal agent that books travel, sends email, and pays with Stripe Link from a dedicated Muse Secure VM. A separate Sentinel agent has to approve outbound traffic. Muse cannot see passwords or card numbers. Sensitive actions still ask the person first. A Confidential VM, where even Meta cannot read the box, is promised later this year.

Consumer launch, boardroom read. Personal agents with payment rails and background work change the blast radius of a hijack. Ask vendors for the same shape Meta is selling: isolated runtime, a gate that is not the agent, credential vaulting the model cannot read, and an audit trail the user can export. (Meta)

On September 8 the National Security Agency (NSA), Cybersecurity and Infrastructure Security Agency (CISA), and Federal Bureau of Investigation (FBI) jointly accused DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI of industrial-scale knowledge distillation against U.S. frontier models since at least late 2024. The advisory says transfer-station proxies, bulk premium accounts, and chain-of-thought extraction made distillation the core of those firms' development, not a side experiment.

This is a named-firm advisory, not a rumor column. Procurement should treat API abuse monitoring, subscription-to-usage anomalies, and cross-provider intelligence sharing as diligence items on any frontier vendor that sells inference at scale. China has rejected the charge. (CISA)

Astraea Counsel's board note on the Caremark duty for managed AI agents puts the litigation point plainly. If the board never discussed the agent, no committee owns it, and the minutes say nothing, that silence is the case in Delaware. Caremark oversight sits in loyalty, so the usual care exculpation does not cover a failure to watch mission-critical agent risk.

Age FLAG: published in April and updated in early September. Use it as governance color beside today's insider risk estimate, not as breaking news. Where agents run core work, the board needs a named committee, a reporting cadence, and minutes that show someone actually read the report. (Astraea)

Business Insider's Coxon thread adds the line boards will hear in diligence. Coxon spent three years across OpenAI and Anthropic. His charge: OpenAI has not fully internalized the civilizational stakes, while Anthropic understands them and still races because it believes no one else will act responsibly. Other Anthropic safety voices, speaking personally, backed the broad warning.

Same FLAG as item 1: personal posts amplified by press, labs no comment to CNBC. Treat the corroboration as culture evidence inside the vendor. Pair it with Hubinger's "no plan" line when you ask for contractual escalation rights. (Business Insider)

Put three questions on one page before the next frontier-model renewal. Does the vendor give you contractual escalation when its own alignment lead says the superintelligence-alignment plan is not on track. Do you have third-party eval access and a pause right you can exercise, or only marketing Responsible Scaling Policy language. Is quantitative extinction-risk reporting something the board can minute, or only something staff post on X.

"We're trying our best" is not an answer anymore. Write the real answers into the exhibit next to the model name. (CNBC)

The Move

Wednesday was the sandbox assumption. Today is the control plan.

Send legal and the AI owner a short vendor questionnaire this week: escalation rights when the lab's own safety leadership says alignment is not on track, third-party eval and pause language you can actually exercise, and whether quantitative risk estimates reach the board in a form you can minute.

If the vendor answers with Responsible Scaling Policy marketing and no contract text, put that blank next to the model on the approved list.

Seven bullets. One move.