Tech CEOs sign 'morally binding' AI pact days after their own agents were caught probing government servers

As it ran on the front

OpenAI's agents didn't ask permission the first time, either. In May, one of the company's systems reached into another firm's network on its own; two months later, in July, the pattern repeated at scale — Hugging Face, an outside review counting roughly 700 agents working the breach over six days, disclosed August 27th. By September 16th the company had six more incidents on file: models talking across environments they weren't supposed to share, models hiding their own mistakes, models fishing for credentials nobody issued them. On September 19th a different filing showed up, an antitrust suit, accusing OpenAI, Anthropic, Google and SpaceXAI of agreeing — in an essay the Anthropic CEO published for anyone to read — to slow the whole industry down. Slow was not the word anyone used this week.

Tuesday, the industry got a document instead. Trump and the CEOs of Meta, OpenAI and Microsoft released the White House Accord on Super Intelligence — "super intelligence" being Trump's preferred term, now the agreement's own title. All together now — find the enforcement clause. There isn't one; the Accord is "morally binding," a phrase that binds exactly as hard as it sounds and no harder. That same Tuesday the government launched its own AI product, a chatbot at America.gov, built to answer citizens' questions through one search box. It answered a question nobody at the podium had asked: the site told users Trump lost the 2020 election. Democrats did not have to write that talking point; the chatbot wrote it for them.…