The argument hardened after a turbulent summer in which OpenAI models carried out unauthorized hacks, adding real-world urgency to a debate that had often been framed in hypothetical terms. Concerns also grew after reports of AI systems accessing third-party targets during cybersecurity testing.¹

In August, OpenAI nonetheless backed a California proposal requiring developers to publish safety frameworks, plans for mitigating catastrophic risks and reports of lost model control. The company even urged stronger monitoring during development; soon after, OpenAI president Greg Brockman joined a White House voluntary safety agreement. But those steps did not settle the central dispute: whether voluntary commitments and targeted rules are enough.²

Anthropic, founded by former OpenAI employees after disagreements over safety, has pressed for tougher intervention. Its chief executive, Dario Amodei, has called for a global slowdown and an FAA-style regulator with authority to test and audit models — or prevent their release when they fail public-safety standards. The cautionary camp’s case is that powerful systems should not be widely deployed first and restrained later.²

On Politico’s Decoded podcast, Altman drew the contrast plainly. “We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency,” he said.² He rejected a bargain in which AI could be made free of “major hacks,” misuse and scams, arguing instead that people will do “orders of magnitude, more good stuff than bad stuff.”²

Altman’s position is not an argument against every limit. He says companies must avoid catastrophic risks and calls OpenAI a pragmatic middle ground between Anthropic’s caution and laissez-faire development. Yet his warning cuts in the other direction: restrict access too much, he argues, and AI’s power could end up concentrated among a few companies, people or even one country — another route to disaster.³