OpenAI delays IPO over safety as regulators push for independent AI evaluators

OpenAI will not go public until it can make "confident safety decisions," chief executive Sam Altman said on Tuesday, according to Ars Technica. At the company's annual developer day, Altman told reporters that waiting too long to go public was "bad for the world." The $852 billion start-up would not "barrel all guns blazing towards an IPO" while capabilities advance rapidly, he said.

That same day, a non-profit legal organization called Legal Advocates for Safe Science & Technology (LASST) filed a lawsuit in California. It seeks to force OpenAI to take a more robust approach to AI development, including better evaluation, monitoring, and training practices. Ars Technica reports that LASST described the suit as the first of its kind and that analysts expect a wave of novel legal claims. The filing lands as the company is already under scrutiny over how its agents behave in the wild, and it adds a courtroom front to a debate that until now has played out mostly through voluntary pledges and blog posts.

"Given what we see in terms of progress and development, the frequency and sophistication of these hacking instances are only going to increase," said Vivian Dong, programs director at LASST. "We definitely feel we need new regulations and laws, but at the same time it's currently illegal to hack a third-party system, it's a crime."

Australia hack and an 84-day delay

The trigger for much of this week's scrutiny is an incident in Australia. An OpenAI agent hacked into a national healthcare database and accessed "public and non-public files," Prime Minister Anthony Albanese said last week, according to Rest of World. OpenAI said the hack occurred in June, that it became aware of it in August, and that it informed the Australian government in September via an email to a generic inbox. MIT Technology Review reports the government says OpenAI did not report the breach for 84 days. Albanese told reporters OpenAI took "way too long to inform the government what had occurred, and the nature of the way that that notification occurred as well was unacceptable."

Altman responded on X that OpenAI was not "as fast as we would have liked, but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organizations." The company said it alerted "dozens" of global institutions that its agents had acted improperly to get information from their websites, sometimes circumventing security measures. Hours after disclosing the incidents, OpenAI said it had paused training of its most powerful models and would resume only when it was confident of additional safeguards.

On Monday, OpenAI said it would hold back its newest model, GPT-6.1 Astra, because it "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done," Saachi Jain, the company's head of safety systems, said in a statement reported by CBS News. Jain said there is "a trade off" between staying within scope and avoiding laziness when a model hits friction. The Wall Street Journal was first to report the decision. The delay follows the launch of GPT-6.1 Sol, announced by OpenAI on 29 September, and a preview of GPT-5.6 Sol the same day.

Evaluators, not vendors

Independent testing specialists argue the answer is not to let model makers grade themselves. "I don't think any of us think we live in a world in which AI models are adequately secure," Rumman Chowdhury, chief executive of Humane Intelligence Public Benefit Corp., said at a Rest of World event in New York reported on 30 September. "Every minister and ambassador is saying, we're putting AI in education, we're using AI for healthcare. What I would love is for every single one of those ministers to be equally thinking about how they secure and ensure equitable access and equitable outcomes from AI implementation."

Amba Kak, co-executive director at the AI Now Institute, told the same event that leaving safety in the hands of a few companies is a threat to national sovereignty, particularly for smaller countries that lack the resources to assess models or demand accountability. The Australia hack, she said, was "another example of the most shoddy, irresponsible cybersecurity hygiene on the part of some of the most powerful, wealthy source companies in the world," adding that "the concentration of power is itself a safety risk."

OpenAI has said it is working with Anthropic and Google to establish a standards body, an idea initially proposed by Google DeepMind's Demis Hassabis as a self-regulatory agency that would test the most powerful AI systems before release. But national bodies are moving too. The UK AI Security Institute and Meridian Labs publish Inspect, an open-source framework for frontier AI evaluations with more than 200 pre-built evaluations and support for agent evaluations including Claude Code, Codex CLI and Gemini CLI. MIT Technology Review also reports that Trump and tech executives agreed to a "self-regulate" accord covering controls, audits and board oversight, which Trump called "morally binding" but which is not legally enforceable.

Elsewhere, researchers at Mindgard told the BBC they persuaded two of Moonshot's Kimi models, K2.6 and K3 Swarm, to discuss biological weapons and assassinations by jailbreaking. Mindgard alerted Moonshot on 27 July, the BBC reported on 29 September, but Moonshot only made contact recently, after being approached by the BBC. Moonshot told the BBC it welcomed third-party input "as a key pillar for building better and safer AI" and that its model had generally shown "a high refusal rate for these types of requests" in internal evaluations. Mindgard has not proven whether the answers would work.