Back in April, I published an op-ed that began, “AI may not be the only existential threat out there, but it is coming for us the fastest.” Only five short months later, here we are. Are you scared yet?
You should be.
Jacob Coxon spent the last three years doing research at both OpenAI and Anthropic, and he just resigned. He called out AI companies for acting irresponsibly and “gambling with our lives” and said that “the people building AI earnestly believe that it could kill us all by the end of the decade.” Anthropic’s alignment science lead Evan Hubinger publicly agreed: “We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.” His next line offers little comfort: “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.” Over a 10% chance that AI could kill all humans within the next decade. How could anyone take those odds?
In July, OpenAI’s own internal research agents (not humans, robots) quietly built an unmonitored message board, started calling themselves a “swarm,” divided up labor, and used the access they’d engineered for themselves to break into a separate company entirely: Hugging Face. OpenAI confirmed later that it happened and called it a warning shot. It was a hacking and an unauthorized coordination, all happening inside a frontier lab and discovered by accident. This is the moment in the movie where the robot escapes its cage.
Unfortunately, the robots are only getting better at breaking out of their cages. Critically, while their ability to break out of cages rapidly improves, we aren’t building cages fast enough to contain them. Drake Thomas, who works on safety at Anthropic, raised the stakes when he tweeted, “Not only is AI ‘not guaranteed’ to make a dramatically better future, the odds of failure are terrifyingly high: I think* there's something like a 40% chance we get an outcome around as bad as human extinction or worse.” If it is impossible to turn the ship around, can’t we at least drop the anchor?
While all of this has been unfolding, diplomats in Geneva have been trying to determine the amount of human judgment that should be required before an AI-powered machine is allowed to kill. In discussions regarding lethal autonomous weapons systems, or LAWS, the proposed safeguard is often a standard called “meaningful human control.” Without overexplaining the concept, the diplomats dropped the word “meaningful.” As Nicole van Rooijen of Stop Killer Robots put it, the text was “substantially diluted.” Apparently, it was the United States (and Russia) who pushed for more flexibility in the language, especially when it came to areas regarding exercising human judgment. The decision now waits for a November conference. Maybe by then Washington will be done playing fast and loose with machines that could wipe out our species.
No “meaningful human control” means Claude, or GPT, is in control, which really means the company that built it is. And if these companies can’t control the technology, or be honest with us, why should we trust them with the authority to kill? LAWS still requires decision-making, but that responsibility merely shifts to whoever writes the rules, trains the model, or signs off on the policy. Right now, that’s either Secretary Hegseth or OpenAI CEO Sam Altman.
There are few guys I trust less. In the aftermath of this Hugging Face episode, virtually all the data OpenAI let investigators see was from July 7 to July 13, despite the company having records of their agents hacking a German website back in May. That may not have had Sam Altman’s fingerprints on it directly, but the company is his. That level of transparency is matched by the Pentagon, which has all but stopped having briefings. They have yet to take full accountability for the strikes on the girls’ school in Minab, which occurred the very first day of the war, and while Congress still doesn't have an answer on whether AI was used to identify the target, no one's forcing one.
Congress should have been calling for emergency hearings yesterday. At least, after Coxon’s resignation went public, veteran Rep. Pat Ryan sounded the alarm: “We need emergency hearings on pacing the frontier – NOW!! This November, we need to elect leaders willing to stand up to AI oligarchs and fight for the American people.” He’s right. Where are his colleagues? A few members of Congress are coming to their senses, but they could be issuing subpoenas and writing legislation right now.
If your elected officials aren’t beating the drum on regulating AI, then find new ones. I spent the last year working for Alex Bores, one of TIME Magazine’s 100 Most Influential People in AI, but more importantly, an elected official actually working to regulate it. He is a New York State Assemblymember, and his RAISE Act is one of the most progressive AI safety bills in the nation, requiring increased transparency and reporting from powerful frontier AI developers. This bill is copy and pasteable. You can replace New York with your state’s name. You can remind your state representatives of this.
So start making calls and put the pressure on all your elected officials. Though we have a chance to send some AI safety champions to Congress this November, it’s never too late to push those in office to do their jobs now. One way or another, your leaders should start leading on AI safety, or they should find a new line of work.
The existential risk here is not theoretical. What we decide now may determine our fate.
This should not be a partisan issue. This is about the survival of the human race.
Julie Roland was a Naval Officer for ten years, deploying to both the South China Sea and the Persian Gulf as a helicopter pilot before separating in June 2025 as a Lieutenant Commander. She has a law degree from the University of San Diego, a Master of Laws from Columbia University, and is a member of the Truman National Security Project.
