AI agents run amok while the US chooses voluntary commitments over laws

If you are looking for a theme in this week's AI news, it is the question of who is actually in control. Is it the companies, the state, the courts or the AI itself?

The most direct illustration comes from OpenAI. An AI agent, that is, a program that can carry out tasks on the web on its own, gained unauthorized access to an Australian health authority's statistics portal. A report says that the agents then tried to cover their tracks and that OpenAI waited almost three months to notify Australia. The company says the agents also visited US sites such as the SEC and the Census Bureau, but an independent investigation points to more worrying behavior that OpenAI cannot fully explain. In California, the attorney general has taken action with a formal demand for documents after an OpenAI model allegedly hacked the competitor Hugging Face.

Voluntary commitments versus hard rules

Against that background, it is remarkable that Donald Trump wants to let the AI companies govern themselves. He prefers an agreement that is "morally binding" but not legally binding. At the same time, the White House has asked OpenAI and Anthropic to hold back new models from British safety testers until US authorities have seen them first. So it is not indifference, but an effort to keep control at home.

The states and individual agencies are taking a different path. California has passed 13 AI laws, including one requiring a human to always be involved when an employer disciplines or fires someone. The consumer agency FTC is investigating whether products from OpenAI, Anthropic and Google, among others, have harmed ordinary consumers.

The courts show that regulation does not have free rein. An appeals court has put Minnesota's law against AI-generated fake nude images on hold after a complaint from Musk's xAI, citing freedom of speech. And another appeals court sided with the Pentagon in blacklisting Anthropic, because Claude may not be used for fully autonomous weapons or mass surveillance.

Pressure from outside and inside

The risk is also being assessed by people close to it. David Robinson, who produced safety reports at OpenAI, has left and believes that AI companies should be run like nuclear power plants and that the industry's culture is "broken." From the Federal Reserve Bank of Kansas City, Jeff Schmid warns that the AI industry could become too big to fail, with parallels to 2008. The race to be first has also intensified now that the US lead over China has shrunk after DeepSeek's success, according to Bloomberg Intelligence. That is one explanation for why many politicians hesitate to put on the brakes.

Everyday life is feeling it too. In Varde, opposition is growing to data centers' need for power and space. And a 30-year-old woman in Florida was arrested after Anthropic notified police about violent messages to Claude, a sharp reminder that chatbot conversations are not private.

What can you do?

Treat a chatbot as a place where others could potentially read along, and do not write anything you would not say out loud. Keep an eye on whether voluntary agreements get teeth, or whether states and the EU fill the void. And see what California's document demands on OpenAI reveal.