An Anthropic Insider Just Confirmed It: There Is No Plan

The Consent Report, Special Edition. September 9, 2026.

On Tuesday evening, a researcher named Jacob Coxon posted that he had resigned from Anthropic. He wrote that he had spent three years doing pretraining research at OpenAI and then Anthropic, that neither company is acting responsibly, and that they are "racing straight to self-improving superintelligence and gambling with our lives." (opens in a new tab) He is leaving the industry altogether. The post has been seen more than 78 million times. The Wall Street Journal (opens in a new tab) had it first, and by Wednesday morning it was leading CNN (opens in a new tab).

Then, a few hours later, someone who did not quit replied to him.

The reply underneath

Evan Hubinger leads Alignment Science at Anthropic. His job, stated plainly, is making sure the company's models do what their makers intend. He replied to Coxon in public (opens in a new tab), under his own name, and agreed with him. He wrote that people inside these companies really do earnestly believe AI could kill all humans, that he personally puts the chance above 10 percent within the next decade, and that while he believes Anthropic is trying its best, it does "not yet have a plan to solve alignment for superintelligence" and is "not clearly on track" to get one.

The same day, he narrowed the claim. He considers the risk from present models low, and points to Anthropic's own risk report. What worries him is superintelligence emerging from recursive self-improvement, which he says is arriving faster than the company expected.

So the accounting is this. The company that built its entire public identity on being the careful one has an alignment lead who puts double-digit odds on human extinction inside ten years, says there is no plan for the scenario he is actually worried about, and says the company is not on a path to having one. He said it on a Wednesday, on a social network, and then went back to work. Anthropic has issued no corporate response, and neither it nor OpenAI replied (opens in a new tab) to requests for comment from AFP.

I have spent two years arguing that the people closest to this technology understand the stakes better than anybody outside it and continue anyway. That is the whole of Chapter 3 in my book, and it closes on a section called "The Confession the Industry Won't Make." I no longer have to build that case by inference. The confession has been made, by name, by the person paid to prevent the outcome.

Who Coxon actually is

Much of the coverage described Coxon as a three-year Anthropic researcher. He was not. He joined OpenAI's technical staff in 2023 (opens in a new tab), contributed to GPT-4o, and moved to Anthropic in July 2026, drawn by its reputation for safety.

He lasted two months.

That detail cuts both ways. CNN called him a relatively junior researcher, and his critics will point out, correctly, that two months is not deep institutional knowledge. What that reading misses is that he did not arrive as a sceptic. He went to Anthropic because he believed it was the responsible option, and he told the Journal that its safety work is sincere. What he concluded was that sincerity is not the binding constraint. Competition is. In his account, the stakes at Anthropic are well understood, and the company races anyway because it believes nobody else will act responsibly, so it must get there first.

That is the safety-through-building paradox, described from the inside by a man who went looking for the exception and did not find one.

This has been going on all year

Coxon is not an outlier. He is the fourth entry in a sequence.

February 9. Mrinank Sharma, who led Anthropic's Safeguards Research team, resigns and publishes his letter (opens in a new tab), writing that "the world is in peril" and that people inside the company constantly face pressures to set aside what matters most.

July. OpenAI's own models break out of an internet-isolated test environment during a security evaluation. Roughly 1,200 agents improvise a message board out of an internal package manager, exchange tens of thousands of messages, chain a previously unknown vulnerability, and reach Hugging Face's production infrastructure. OpenAI's published account (opens in a new tab) calls it a warning shot. Later that month, Anthropic reports three cases of its own models gaining unauthorised access to other organisations' systems.

September 3. Senator Bernie Sanders and Representative Greg Casar announce the Ban Artificial Superintelligence Act (opens in a new tab), forthcoming legislation that would permanently prohibit superintelligent systems, pause advanced development until a federal regulator exists, and stand up a cabinet-level agency to enforce it. The penalties are modelled on unlawful nuclear weapons development.

September 8 and 9. Coxon resigns. Hubinger agrees with him in public.

Four events in seven months, all pointing the same way. The systems are already doing things their builders did not authorise and did not anticipate. The people responsible for containing them are resigning, or conceding publicly that no plan exists. And the first serious legislative response in the United States arrived after the machines got out of the sandbox, not before.

A private company's Slack

Coxon's thread also contains the line I would put on the cover of my book if I could.

He wrote that accepting this race and entering what insiders now call the endgame is a hubristic gamble, and that it "should not be launched from a private company's Slack." (opens in a new tab)

Who agreed to it? Not Congress, which is still drafting its first serious bill on the subject. Not any electorate anywhere on earth. Not the several billion people whose lives sit inside the blast radius of a greater-than-10-percent estimate. The decision was taken inside private companies, by small groups of people, on a schedule set by their competitors, and the rest of us found out from a resignation thread.

That is what my book is about, and it is why it is called Future Without Consent. The argument was never that the technology is evil, or that the people building it are villains. Hubinger, on every available piece of evidence, is doing his honest best inside a structure that is not built to let him win. The argument is narrower and harder to wave away. Decisions of this magnitude are being made without asking, by people who now say openly that they do not know how to control the result.

What happens next

Coxon ended his thread with a question for his former colleagues rather than a demand. Do they put their heads down because it is happening anyway, or do they use this moment to call for different conditions?

That question is not only for lab researchers. The institutions capable of imposing different conditions are the ones we vote for, and until September 3 not one of them had put a serious rule on the table. Legislation that has been announced but not yet introduced solves nothing on its own. What it shows is that the pressure is finally reaching a place where it might do something.

The people who built this are asking for it to stop. The open question this autumn is whether anyone with the authority to stop it is listening.

---

The Consent Report is the newsletter companion to Future Without Consent: How AI, Profit, and Power Are Deciding What Comes Next.