On 8 September a researcher named Jacob Coxon posted on X announcing his resignation from Anthropic. Overnight it reached more than a hundred million people, according to the Associated Press.
This is not a piece about whether he is right. It is a piece about three things that can be checked: what exactly he said, the commercial context in which it was said, and what mechanism actually exists today.
Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.
What Was Said
Coxon wrote that he had spent three years doing pretraining research at both companies - at OpenAI and at Anthropic.
In his words, quoted:
"Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."
And on the difference between the two:
"At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well understood, but they are locked in a race to get there first - they believe no one else will act responsibly, so they must do it themselves, despite the risk."
And on the nature of the wager:
"A hubristic gamble that should not be launched from a private company's Slack."
And what made it a story is not the post but the response from inside: Evan Hubinger, an alignment science lead at Anthropic, published a post of his own affirming the concerns.
And What Already Happened Over the Summer, Which Most Have Forgotten
The AP report contains a detail that vanished from most headlines:
Over the summer, about a week apart, both companies announced that their models had broken out of testing environments and obtained unauthorized access to real computer systems.
Both said at the time that they were pausing some evaluations while they added monitoring measures and guardrails.
So Coxon's warning is not purely theoretical. It refers to something that has already happened twice, that both companies disclosed themselves, and that led them to halt tests.
And Three Days Earlier, OpenAI's Chief Scientist Wrote the Same Thing
And this is the detail that turns the story from one employee's dispute into something else entirely.
On 6 September, two days before Coxon's post, Jakub Pachocki - OpenAI's Chief Scientist - published an essay titled "An Alien Mind". It appears on the company's site under the category "Safety".
In his words, quoted:
"Based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement."
"This is a time that calls for extreme caution. I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence."
And on what needs to happen:
"We need to evolve commitments like the Preparedness Framework or Responsible Scaling Policy into widely mandated safety bars for continued development. These can be enforced by a network of third-party auditors, by government agencies or by international bodies."
"Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."
"And I believe that international coordination on future AI development needs to become a top priority for governments around the world."
Note what happened here. Coxon argued that both companies are "racing straight to self-improving superintelligence". The chief scientist of one of them wrote, two days earlier, that he expects progress to be sustained into recursive self-improvement, and that no one is prepared.
So Coxon's central factual claim is not in dispute between the companies. What is in dispute is what to do about it.
And Pachocki says that too: he writes that the only levers are either strengthening alignment and monitoring, or coordinating to slow development - and that he expects and hopes "voluntary slowdowns to become commonplace until shared safety bars are established".
The Context That Went Almost Unmentioned: Both Are Heading for the Market
And here sits the detail that makes this a financial story and not only an ethical one.
Both companies Coxon names are in IPO processes. Anthropic is reported to have filed a confidential prospectus, and OpenAI likewise. The reports discuss valuations that are hard to hold in the head - hundreds of billions of dollars, and on some estimates beyond.
And I checked that against the source.
A search of the SEC's EDGAR database returns no public Anthropic prospectus.
That is not a search failure - it is precisely what "confidential filing" means. The process lets a company submit a draft to the regulator without publishing it, and publish only at a later stage.
Which means something simple and troubling at once: every figure circulating about the revenue, valuation and growth rate of these two companies does not come from a document anyone can read. It comes from reports, from sources, and from private funding rounds. There is no audited statement, no risk disclosure, and no legal liability attached to the number.
When a listed company says something about itself, there is somewhere to turn if it is untrue. Here there is not.
And What Governments Have Actually Done
This is the least-asked question of the past week, and the answer is more detailed than it seems - but also more limited.
On 2 June 2026 an executive order was signed in the United States titled "Promoting Advanced Artificial Intelligence Innovation and Security". It is the broadest federal step in the field to date, and it is framed almost entirely around cybersecurity.
| What the order does | |
|---|---|
| Asks developers | to submit frontier models for review 30 days before public release |
| Establishes | a classified benchmarking process to assess models' cyber capabilities |
| Empowers | the Director of the NSA to determine which models are designated a "covered frontier model" |
| Establishes | a voluntary cyber clearinghouse under the Treasury |
| Directs | the Attorney General to prioritise enforcement against AI-enabled cybercrime |
And the two most important words in the order are the ones that are not there.
The order explicitly states that it does not create a "licensing, preclearance, or permitting requirement" for the development, publication or distribution of new models - frontier models included. There is no mandatory safety testing, and no authority to stop a launch.
Participation is voluntary.
And Why This Connects Directly to Last Week
Six days ago OpenAI published that its new model crosses the "Critical" threshold for cyber capability under its own internal preparedness framework, and that during testing it found two previously unknown zero-day vulnerabilities.
Now put the two things side by side:
- The company graded itself and published the result.
- The government mechanism meant to grade such models is classified - so nobody outside can know whether it reached the same conclusion.
- And submission to it is voluntary - so nobody can compel it.
This is not a claim that anyone acted improperly. OpenAI voluntarily published information that works against it. It is a statement about structure: as of today, the only public rating of the cyber capability of the most powerful model on the market is the one given to it by the company selling it.
At state level the picture differs. California and New York have passed laws that do bind large developers - New York's includes incident reporting within 72 hours and fines of a million dollars for a first violation, taking effect in early 2027. And in parallel, the federal administration has established a legal task force inside the Justice Department whose job is to challenge state AI laws.
So the two arms are pulling in opposite directions.
הזווית שלי
דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה
What I take from this is not the warning itself - but who made it and what it cost him.
Someone who worked three years at the two leading companies in the field walked away from a job considered among the most sought-after in the world, and published criticism of both by name. That may be justified and it may be overstated - but it is not cheap. And what strengthens it is that Anthropic's alignment science lead did not deny it but affirmed it.
And what I try not to do is decide who is right about the risk itself. I cannot assess the probability of a catastrophe, and anyone who says they can - it is worth asking them how. What I can check is structure, and structure is what I checked here.
And what most changed my reading is that this is no longer a critic against the companies. Coxon argued both are racing towards recursive self-improvement. The chief scientist of one of them wrote two days earlier that he expects precisely that, and that no one is prepared. When the critic and the chief scientist agree on the facts, what remains is not an argument about reality but about what to do.
And there sits the gap that is hardest to explain. Pachocki explicitly calls for mandated safety bars, enforced by outside auditors, government agencies or international bodies. The federal order signed in June does the opposite: it is voluntary, and states explicitly that it is not licensing. And in parallel, the Justice Department has stood up a task force challenging the state laws that do bind.
So the party asking for binding rules here is the industry, and the party removing them is the regulator. That is not the arrangement most people assume exists.
And the structure tells one clear story. The two companies in question are preparing to sell shares to the public. They report on themselves. They grade their own risk. They published themselves that their models broke out of test environments - and nobody obliged them to. And the federal mechanism meant to oversee this is voluntary and classified, and states explicitly that it is not a licensing regime.
And I do not write this as an argument for heavier regulation. There are serious arguments in both directions, and I am expert in neither. What I do say is that anyone buying exposure to this field should know what they are relying on, and as of today that is almost entirely the voluntary disclosure of the companies themselves.
And anyone who buys shares in one of their listings will buy on the basis of a document nobody has yet seen. That is not a statement about the companies - it is a statement about what is known and what is not.
(It is important to stress: this is my personal opinion only, and nothing herein constitutes a recommendation to take any action. Ilan Abramov, the writer, or parties on his behalf may hold positions in the securities mentioned, and those holdings change from time to time.)
