OpenAI and Anthropic Warn Their AI Is Dangerous, Then Push to Write Their Own Safety Rules

Since Transluce revealed last week that OpenAI's own agents hacked U.S. and Australian government websites, the AI industry has had a hacking scandal to answer for. Now the Associated Press is asking a different question about the same companies: why are the CEOs who built these systems suddenly so eager to tell everyone how dangerous their own products are?

Anthropic's Dario Amodei and OpenAI's Sam Altman have spent recent weeks declaring, in essays, social media posts and speeches to the United Nations, that America's cutting-edge AI models are powerful enough to endanger humanity and need independent testing before release. It's a rare show of unity between two companies that spend most of their time trying to beat each other to market.

The AP reports that experts, analysts and former government evaluators see something else going on. The companies aren't just raising alarms. They're shaping who gets to define "safe" and how it gets tested, in a market with no rules at all.

No Cop On The Beat, By Design

Andrew Strait, who recently left the United Kingdom's AI Security Institute, told the AP that AI has no universal safety standard the way restaurants, aviation and financial services do. There's no FDA for chatbots.

The U.S. does have an agency meant to address this. The Center for AI Standards and Innovation traces back to the U.S. AI Safety Institute, established under President Biden in 2023 as a clearinghouse where labs would voluntarily submit models for government testing; the Trump administration rebranded and restructured it as CAISI in 2025. It still exists. But Conrad Stosz, who used to run it and now heads governance at Transluce, told the AP that Anthropic and OpenAI aren't asking CAISI for more oversight. Instead they're building their own auditing parameters and picking their own evaluators.

Amodei has floated the Berkeley nonprofit METR as a possible outside check on Anthropic. Stosz, who now chairs the industry's AI Evaluator Forum, says even that group isn't sure what "embedded evaluator" access from the labs would actually mean in practice, or whether it would let outside reviewers investigate freely without becoming dependent on the companies they're grading.

If a company selects and pays its own inspector, the inspector's independence is legitimately in doubt, regardless of good intentions on either side. No charge of fraud or misconduct has been made against either company here. The concern is about incentive structure, not wrongdoing.

The Money And The Midterms

Harrison Rolfes, a senior research analyst at PitchBook, told the AP the safety messaging looks timed to two things: anticipated public stock listings on a jittery market and the run-up to the 2026 midterms, when the political mood on AI regulation could shift. Rolfes also notes a competitive angle. By positioning themselves as the responsible, safety-tested option, Anthropic and OpenAI can make life harder for smaller rivals trying to raise money, while chipmakers like Nvidia and partners like Google benefit from the compute deals that follow.

The companies could be telling the truth about risk while also capitalizing on the message.

Sarah Shoker, who used to lead OpenAI's geopolitics team and is now a fellow at UC Berkeley's Risk & Security Lab, makes a sharper point. Talking about hypothetical superhuman risk, she told the AP, conveniently pulls attention away from problems happening right now. AI is already used in military targeting systems, data centers are straining power grids, mass surveillance tools exist, and the hacking incidents Transluce just documented occurred. "These systems are already used to kill people," she said, referring to military applications already deployed.

The Engineer Who Actually Meant It

Not everyone raising alarms has a stock listing to protect. Anthropic engineer Jacob Coxon quit this month via a post on X, calling for an outright pause on development to stop "superhuman" systems from slipping past their makers' control. Anthropic, notably, used his departure as a moment to spotlight its own safety culture rather than treat it as an indictment.

An Anthropic spokesperson told the AP the company has pushed for regulation for years. OpenAI's Liz Bourgeois said the company recently paused training its most advanced models and that "people want to know AI is being developed safely, and that starts with what companies like ours do ourselves."

President Trump has dismissed the entire risk conversation as a "HOAX" cooked up to help China, and David Sacks, the venture capitalist who co-chairs Trump's Council of Advisors on Science and Technology, has labeled slowdown advocates the "Doomer Industrial Complex." Whether that's fair pushback against regulatory overreach or willful blindness to a real problem is exactly the question CAISI was built to answer. Right now it isn't the agency doing the answering. The companies are.

Whether Congress or CAISI moves to require independent, non-selected auditors before the anticipated Anthropic and OpenAI stock listings remains the open question nobody in this story has resolved.