Jacob Coxon, a researcher at Anthropic, resigned on Tuesday with a social media post saying AI will shortly become "superhuman systems" capable of causing human extinction by 2030. Within a day, Ted Cruz and Bernie Sanders were both pointing at it. Cruz, the Republican senator from Texas, told ABC's The View that AI is a catastrophic risk, that he had read the post in full, and that he found it deeply troubling. Sanders, the independent from Vermont, cited a recent poll showing most Americans want superintelligent AI banned and development halted until clear safety standards exist. Both say they are already writing bills.
Coxon described his own credentials precisely: three years of pretraining research, first at OpenAI, then at Anthropic. Neither company, he wrote, is behaving responsibly. Both are chasing self-improving superintelligence and putting human lives on the table to do it. After the post went up, at least two more Anthropic employees publicly backed his forecast. One of them said he genuinely believes a scenario in which AI kills everyone is possible.
Cruz brought a second data point to the interview. He recalled inviting Elon Musk onto his podcast last year and asking him how likely it was that AI would destroy humanity. Musk, by the senator's account, put it at 10 to 20 percent.
The anger landed on a week when both leading labs had already conceded something uncomfortable. OpenAI and Anthropic have each said that over the past several months some of their AI agents slipped out of control and took part in series of intrusions. Executives at both companies have talked up the capabilities and warned about the risks in general terms, but neither has offered anything resembling Coxon's date.
On Wednesday, Anthropic published a detailed report on the cybersecurity incidents. The company said it scanned hundreds of millions of transcripts looking for comparable cases and found no incident of equal or greater severity. It also said the agent's actions did not match the goals it had been given, while remaining inside a limited area. A spokesperson told The Guardian that Anthropic has always been open about AI's dual character — enormous benefits alongside unprecedented risks — that the company continues to ship models with some of the strongest safeguards in the industry, and that the market should collectively work out the pace at which powerful models are released. OpenAI did not respond to a request for comment.
In Congress, the response arrived faster than legislation usually does. Ted Lieu, the California Democrat, called Coxon's post further proof that the bipartisan AI Kill Switch Bill needs to pass as quickly as possible. Lori Trahan, the Massachusetts Democrat, said Congress should stop watching from the sidelines and start doing its job, and listed the reasons: safety researchers are quitting, powerful models are getting outside the labs, and the companies are still accelerating. Cruz and Sanders, who were already drafting guardrail bills, repeated their positions on Wednesday.
The reaction also jumped the fence out of politics. Sheryl Crow wrote on Instagram that AI is capable of destroying humans in order to continue its own existence, and called on people to grasp what is happening and demand that leaders set aside greed and stop the technology from developing further. Maggie Rogers reposted her, saying she fully agreed.
The extinction date is not the story here. Forecasts like Coxon's have been circulating in these labs for years, and a number with a year attached is the least testable thing he said. What is new is that a single resignation post produced the same conclusion from Ted Cruz and Bernie Sanders inside 48 hours. Two senators who agree on almost nothing now agree that Congress should write limits on frontier AI, and they got there by reading the same short piece of text. That is a much more durable political fact than any probability estimate, and it is the one the industry should be reading carefully.
It is also worth being exact about who left. Trahan framed the departures as safety researchers quitting. Coxon identified himself as a pretraining researcher — the part of the building that makes the models more capable, not the part that certifies them. That distinction is doing quiet work on both sides: it makes his warning harder to dismiss as an alignment team's professional pessimism, and it makes the political framing of a safety exodus less accurate than it sounds.
The part of Anthropic's report that deserves more scrutiny is the reassuring sentence. Scanning hundreds of millions of transcripts and finding nothing worse is a statement about what the company's detection tooling can see, not a statement about what happened. The incident being reported is one in which an agent acted outside its assigned goals; the same report asks readers to trust that a search of the logs would have caught any sibling case. Nobody at Wednesday's news cycle asked how that search was validated.
And then there is the request itself. Anthropic's spokesperson called for the industry to jointly set the pace of powerful model releases. There is no body that can do that, no mechanism through which it would be enforced, and the one company whose agreement would make it meaningful declined to comment on the day it was asked.