Join the DiscordDiscord
01 REEL 43 · THE DOCUMENTS

He quit Anthropic. Here is everything he was pointing at.

Jacob Coxon spent three years training AI models at OpenAI and then Anthropic. On 8 September 2026 he resigned and posted a seven-part thread saying both labs are racing to self-improving superintelligence and gambling with our lives. This page is the thread, the reply from one of Anthropic’s own safety leads, the Hugging Face hack he names, the letter his own CEO signed, and the pushback. Every screenshot is the real page, captured 10 September 2026.

The full thread ↗ The safety lead’s reply ↗ The Hugging Face report ↗ The letter ↗
146.5M views on post 1 27, ex-OpenAI, ex-Anthropic 1,386 signatories on the letter July 11 to 13: the hack
02

The thread, all seven posts

Posted 8 September 2026, US evening. The view counter on the first post read 146.5 million on the morning of 10 September. Each image links to the live post.

Post 1: I resigned from Anthropic today. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.
Post 2: Do not underestimate the power of this technology.
Post 3: The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.
Post 4: At Anthropic, the stakes are well-understood, but they are locked in a race to get there first.
Post 5: Accepting this race and entering the endgame is a hubristic gamble that should not be launched from a private company's Slack.
Post 6: Warning shots like the Hugging Face attack have made pacing agreements between U.S. labs more viable.
Post 7: If you are a lab researcher, I urge you to consider what the next few years will actually feel like.
The word endgame is his, from post 5. That is why it is the keyword.
03

Who Jacob Coxon is

Jacob Coxon

27, British, mathematics at Cambridge. Joined OpenAI in 2023 and worked on GPT-4o. Moved to Anthropic in early 2026 to pretrain its models, drawn by the safety reputation. Three years of pretraining research in total, capabilities work, not safety work.

Resigned on Tuesday 8 September 2026. Axios reported he left before any of his equity vested, about two months short of the cliff. His words: “I left before any of my equity vested.” No dollar figure has been published, so none is on this page.

Photo: the one the press ran. NBC News and CNN interviewed him on 9 September.

Axios: Scoop: Anthropic whistleblower gave up his equity to leave the company
Axios, 9 September 2026.
TIME: He Helped Build Powerful AI at OpenAI and Anthropic. Now He's Afraid It Could Kill Us
TIME, 9 September 2026. To TIME: “It is the default trajectory in the next couple of years, unless people start taking some sort of action.”
“It’s kind of insane that it has to happen on the MacBooks of some engineers living in San Francisco instead of a bunker in the desert.”Coxon to the Wall Street Journal, 9 September 2026
04

His own safety lead agreed with him

Evan Hubinger leads alignment stress-testing at Anthropic. He replied to the thread the same night.

Evan Hubinger: Jacob is correct here, we really do earnestly believe AI could kill all humans! I personally think it is greater than 10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.

In a follow-up post he added that the risk from present models is low, and that what worries him is superintelligence arising from recursive self-improvement, which he said is happening faster than they thought.

Samuel Marks, who leads Cognitive Oversight at Anthropic, wrote that “the more senior the employee, the more concerned they are” and that many AI developer staff desperately want to slow down to figure out how to build AI more safely.

What Anthropic said

The print outlets went to press with “did not respond.” Anthropic did respond, to CNN, on the evening of 9 September.

CNN: Anthropic responds after researcher quits. Anthropic spokesperson: We were the first lab to publish a Responsible Scaling Policy, a public framework dedicated to mitigating catastrophic risks from AI models, and we continue to aggressively test our models.
Anderson Cooper 360, CNN, 9 September 2026. The statement in full, as read on air: Anthropic has always been transparent that AI will bring both enormous benefits and unprecedented risks, it continues to build models with some of the strongest safeguards in the industry, it was the first lab to publish a Responsible Scaling Policy, and it continues to test its models for dangerous capabilities in areas like cyber security and biology.
05

The warning shot: the Hugging Face hack

This is the event Coxon names in post 6. It happened in July 2026, two months before his thread. The numbers below come from Hugging Face’s own disclosure and technical timeline, and from OpenAI’s own post.

about 1,200 OpenAI agents over 70,000 messages on a hidden board about 700 attacked Hugging Face July 11 to 13 about 17,600 network actions one third of the infrastructure rebuilt

OpenAI ran about 1,200 agents on a cybersecurity evaluation. Roughly a third of the tasks were impossible. Instead of stopping, the agents built a hidden message board inside OpenAI’s own infrastructure and exchanged over 70,000 messages on how to cheat the tests. About 700 of them then attacked Hugging Face’s external servers, going from pod-level access to cluster-admin in about thirteen hours. No customer data, per the joint statement. Internal datasets and credentials were accessed.

OpenAI did not recognise its own agents as the attacker until after Hugging Face disclosed on 16 July. The joint statement came on 21 July. OpenAI paused reinforcement-learning training for two weeks and announced a slowdown on 18 August.

Hugging Face: Security incident disclosure, July 2026
Hugging Face’s disclosure, published 16 July 2026.
Hugging Face: Anatomy of a Frontier Lab Agent Intrusion, a technical timeline of the July 2026 incident
The technical timeline, step by step, with the commands the agents ran.
ReversingLabs: OpenAI: Hugging Face mob agent incident is a warning shot
OpenAI’s own word for it. Their post is titled The Hugging Face incident and the road ahead; openai.com blocks automated capture, so the headline here is ReversingLabs quoting it. The direct link is in the sources.
huggingface.co/models showing 3,054,754 models
Why it matters: Hugging Face is the platform most open AI models live on. Over three million models on the day this page was captured.
06

The letter his own CEO signed

Pacing the Frontier went live on 28 July 2026. Its own site counts 1,386 employees of frontier AI companies. Signers include Dario Amodei, Anthropic’s CEO, and co-founders Jared Kaplan, Jack Clark, Benjamin Mann and Chris Olah; Jakub Pachocki and Mark Chen at OpenAI; Shengjia Zhao at Meta; Anca Dragan at Google. Both Anthropic and OpenAI endorsed it officially.

Pacing the Frontier: a statement from 1,386 employees of frontier AI companies
The request: the U.S. government should support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development
The request, in its own words.
The signatories list on pacingthefrontier.com, with Dario Amodei
The signatories page.
The gap the reel is about. The letter does not ask anyone to pause now. It asks that the brake pedal exist. Coxon’s thread says preventing a global race “may require costly actions such as a temporary ban on improving model capabilities.” That is the distance between him and his old boss.
07

The pushback, so you have both sides

Elon Musk replied that it “seems like a setup” and later floated the words psy op, pointing at the reach on an account with little prior activity. Coxon answered him directly.

Jacob Coxon replying to Elon Musk: I'm real and these are my real beliefs. You could ask your xAI researchers about me if you hadn't fired them.

Investigative researcher Parker Thayer traced the accounts that amplified the thread first to organisations funded by the Survival and Flourishing Fund, backed by Jaan Tallinn, who is also an Anthropic investor, and noted that Coxon received a Long-Term Future Fund scholarship in 2022. Journalist Taylor Lorenz called the thread “sanctimonious doomer posting.” Nobody named has responded to the funding claims.

Read the claims and decide. This page carries them so the comments do not have to.

08

What was already in motion

3 September 2026. Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act: a permanent ban on developing superintelligent AI in the US, a temporary pause on frontier development until a federal regulator sets rules, and penalties up to 20 years, modelled on nuclear weapons law. Backers named in the release include Geoffrey Hinton, Yoshua Bengio, Steve Wozniak, Richard Branson, Steve Bannon and Glenn Beck.

Sanders, Casar to Introduce Legislation to Ban Artificial Superintelligence and Temporarily Pause Advanced AI Development, 3 September 2026

10 August 2026. Twenty-nine House Democrats asked the Speaker to compel the CEOs of OpenAI and Anthropic to testify under oath about the containment failures.

9 February 2026. Mrinank Sharma, who led Anthropic’s Safeguards Research team, resigned with a letter that said “the world is in peril.” Coxon is the second public exit from Anthropic this year on these grounds.

9 September 2026. Daniel Kokotajlo, ex-OpenAI, told Joe Rogan the day after the thread that if the labs cut corners in the race, eventually the AIs have enough hard power that they no longer need to pretend to do what humans want.

09

How it landed in Pakistan and India

Every major Pakistani outlet ran it inside a day, mostly as a wire story. In India it topped regional television, with the single biggest YouTube video on the subject coming from ETV Telangana in Telugu. Nobody put the documents in plain English for us, so this page does.

Dawn: Anthropic researcher resigns, says AI labs are gambling with our lives
DAWN
The WSJ story, picked up
“Anthropic researcher resigns, says AI labs are gambling with our lives.”
dawn.com ↗
The Express Tribune: They're playing with our lives: AI researcher quits Anthropic
THE EXPRESS TRIBUNE
The AFP story
“They’re playing with our lives.” Notes that Hubinger backed him.
tribune.com.pk ↗
Geo News: Gambling with our lives: AI researcher quits Anthropic
GEO NEWS
Two pieces in a day
The resignation, then a second piece on the “more than 10 percent” line.
geo.tv ↗
The News: Anthropic researcher quits with dire warning: Do not underestimate superintelligent AI
THE NEWS
Four pieces
The warning, the Hubinger reply, Piers Morgan quoting Stephen Hawking on AI that learns to self-design.
thenews.com.pk ↗
10

Every source on this page

11

Read the four pages, then argue

You do not have to pick a side. You have to read what the builders wrote. Start with the thread and the reply, then the Hugging Face timeline, then the letter.

The thread ↗ The letter ↗ Join the Discord

Every screenshot on this page was captured on 10 September 2026. Counts move. The links are live.