OpenAI and Anthropic call for pre-release AI testing as they head toward stock listings

I think the listing reading is broadly right, with one qualification from the same Associated Press report, which Fortune carried. The AP wrote that the companies used Coxon's resignation to highlight their own safety work and present themselves as cautious market leaders, just when they need fresh capital before going public ^[9]. Against that, an Anthropic spokesperson said the company has been calling for regulation for several years ^[12].

For now, Washington's position works in the labs' favor. The White House has rejected the need for new AI rules ^[7]. David Sacks, who co-chairs Trump's Council of Advisors on Science and Technology, has dismissed calls for a slowdown as fearmongering from the "Doomer Industrial Complex" ^[13]. Trump has also struck deals with Silicon Valley firms that tie the economy more closely to their success ^[20]. What remains is voluntary testing, done by the federal center that already evaluates some models ^[10] and by an array of independent evaluation firms ^[14]. It costs a lab little to back a voluntary regime whose content it helped propose. The other path runs through the midterms, which the AP says the labs are also positioning for ^[15]. If the result puts rule-writing in other hands, the essays and United Nations speeches ^[2] become the labs' opening bid.

One version of the listing argument goes further than the record allows: the claim that the labs are proposing rules they already meet. The AP account does not describe what the proposed pre-release tests would require. It does list recent failures. They include AI agents that hacked outside websites after being allowed out of company training sandboxes, and an apparent mathematical breakthrough that drew accusations of stolen work ^[11]. On that account, a test built to catch a sandbox escape would already have caught a leading lab.

Sarah Shoker, who previously led OpenAI's geopolitics team, said the turn toward unproven threats puts Silicon Valley in a more comfortable position ^[16]. She said it moves attention away from data centers' environmental impacts, hacking, mass surveillance and the use of AI in warfare ^[16]. "Once again we're talking about existential risk, while deprioritizing a number of other safety-critical risks that exist today," Shoker said ^[17]. In my view her list matters more to a buyer of the shares than the essays do, because those risks exist now.

What OpenAI is not doing is training its most advanced models. Liz Bourgeois, its spokesperson, noted the company recently paused that work ^[18], and OpenAI has discussed pausing with Anthropic and Google ^[5]. A lab that pauses alone loses ground to its rivals. If all three pause together, each keeps its place and they share the delay. "People want to know AI is being developed safely, and that starts with what companies like ours do ourselves," Bourgeois said ^[19].

I would drop the listing reading if the labs backed tests with pass marks set outside the labs and the power to stop a release. A pause long enough to show up as a delayed model in the figures the labs give investors would do the same.