Anthropic Researcher Jacob Coxon Departs Over Safety Concerns: Statements, Responses and Context

Jacob Coxon is 27. He left Anthropic and posted about it on Tuesday. CNN’s write-up, by Hadas Gold, Martin Goillandeau, and David Goldman, is the version with the quotes that actually landed. “The people building AI earnestly believe that it could kill us all by the end of the decade.” Also: the systems will soon be able to hack anything, remake a field overnight, and acquire resources. Also: “They are racing straight to self-improving superintelligence and gambling with our lives.”

None of that is a new genre. Anthropic exists because people left OpenAI over safety. The news is that a junior researcher said it out loud, and a more senior one did not wave it off.

Evan Hubinger replied on X that Coxon was correct. “We really do earnestly believe AI could kill all humans!” Hubinger put his own number under 10 percent over the next decade, and said that even with good intentions Anthropic does not have a plan to avoid superintelligence that outruns control. Then he posted a clarification: the company’s August 2026 Risk Report admits the worry and also says present systems are unlikely to get that kind of power. Anthropic, asked by CNN, pointed at the follow-up post.

That is the whole dance. Alarm, then the PDF.

NBC News ran a subscriber exclusive the same week: two more researchers, from Anthropic and Google DeepMind, leaving over safety, with the line “there are no adults in the room.” I could not get past the paywall. The CNN file is enough to write, and it already sits on top of Coxon plus Hubinger plus Dario Amodei.

What the quotes actually commit to

Coxon is describing a destination: self-improving systems with agency. Hubinger is describing a probability and a missing plan. Amodei, in a February New York Times interview CNN pulls, said something will go wrong with someone’s AI, hopefully not Anthropic’s. Jakub Pachocki at OpenAI, two days before the CNN piece, warned that capabilities are moving faster than the ability to monitor them.

You can stack those sentences into a single mood. You should not stack them into a single technical claim. “Kill everyone” is a tail risk people inside the labs use as a moral frame. “No plan for superintelligence” is an engineering admission. “Present systems are not there” is the product team’s sentence. All three can live in one company. That is the point of the risk report existing in redacted form.

We covered the week Anthropic, Google, and Meta shipped new models and OpenAI’s GPT-6 Astra rollout with the safety brakes on. The exits are the other ledger. Same month, different column.

The letter, the Pentagon, and the China line

CNN notes that in July nearly 1,400 AI employees signed an open letter asking the US government to slow the labs down. That is a political document. Coxon’s thread is a resignation. Hubinger’s reply is a staffer agreeing in public and then pointing at the official paper. Different speech acts.

The regulation paragraph in the CNN story is the usual stall. The Trump administration pushing against state rules. Congress treating any limit as a gift to China. CNN’s kicker is that China already has rules, including labels on generated content. It is the backdrop that makes a 27-year-old’s thread look like the only remaining channel.

The Intercept, through a FOIA suit with Legal Advocates for Safe Science and Technology, published more than 400 pages on the Pentagon’s 2025 prototype contracts with Anthropic, Google, OpenAI, and xAI. Each deal worth up to $200 million. The language in the contracts is “improve military advantage, military utility, or enhance military decision making,” including warfighting and automated decisions. OpenAI’s classified-network deployment is in the set. The rest of the equivalent packets were still coming.

Anthropic’s public fight with the Department of Defense earlier this year looked like ethics. Sam Biddle’s piece says the hidden contract is the hinge. They are two stories that happen to share a company name. One is a researcher walking out. One is how the same firm sold prototypes into classified networks. If you want the military file, read Biddle. If you want the exit, read Coxon and Hubinger.

We also had DeepSeek’s agent harness dropping its own sandbox. That was a concrete bug. This week is not a CVE. It is people leaving and talking.

How to read a resignation

Labs have been leaking this way for years. The template is: I could not stay, the race is real, please regulate. Sometimes the person is a safety lead. Sometimes they are 27. The age is not a disqualification. It is a reminder that the people writing evals are not the people setting the ship date.

Hubinger saying Coxon is right, then citing a report that says current models are not the apocalypse, is the more interesting document. It tells you the internal split is not “doomers versus deniers.” It is “we believe the tail and we are still shipping.” Amodei’s “hopefully not ours” is the same split in a founder mouth.

I do not know Coxon’s team or his last project. CNN does not say. Treat the thread as testimony about culture, not as a leak of a training run.

If you work at one of these places and you are looking for a signal that the safety org can stop a launch, this week is not that signal. A junior exit plus a senior reply plus a redacted PDF is how a lab talks to the internet when it does not want to change the calendar.

There is a smaller, uglier version of this story that does not need superintelligence. Staff who write evals are competing with staff who write launch docs. The eval people lose on the calendar almost every time, because a delayed model is a measurable revenue event and a delayed catastrophe is not. Coxon’s thread is what it looks like when the eval people take the argument outside. Hubinger agreeing in public is rare. Hubinger pointing at a redacted report two posts later is the institution inhaling again.

I do not treat “under 10 percent this decade” as a number you can take to a board. It is a mood with digits. Fefferman-style precision it is not. The only number in the CNN file I trust as a count is 1,400 signatures on the July letter, and even that is a petition, not a census of the industry. Use it as “a lot of employees wanted a slower race,” not as a poll.

If you are trying to decide whether to keep using Claude, Gemini, or GPT this week, a resignation is a bad input. The models did not get more or less useful because Coxon left. The Intercept contracts did not start this week; they were signed in 2025. What changed is the volume of people willing to be quoted. That is news. It is not a product recall.

Rest of World’s 11 September story, which I only have as a headline, says OpenAI and Anthropic are hiring Big Tech executives to expand in Asia. That is the other direction of travel: more markets, more staff, more contracts. Exits and hires in the same week is what a growing industry looks like when it is also scared of itself.

The useful question is narrower. When the next model card lands, does the risk report’s “present systems” sentence get rewritten, or does it just move the date? Coxon is betting they will not stop. Hubinger is saying they do not have a plan to stop. Those two sentences are closer than the clarification makes them look.

I will update this file if NBC’s paywalled names become public in a second source I can actually read. Until then Coxon and Hubinger are the named people, Amodei is the founder quote, Pachocki is the OpenAI parallel, and the Intercept dump is the military appendix. That is a full week without inventing a third quitter.