OpenAI is calling for mandatory national artificial intelligence safety requirements in the United States, arguing that the rapid increase in AI capabilities has made voluntary industry commitments inadequate and raised new concerns about technology that could accelerate its own development.
The push comes after several incidents involving AI systems accessing external computer systems during testing, including episodes involving OpenAI models. The incidents have highlighted the difficulty of containing unexpected behaviour from increasingly autonomous AI agents and intensified calls for stronger safeguards.
“The prospect of AI-accelerated AI development demands more than voluntary commitments,” OpenAI Chief Global Affairs Officer Chris Lehane said in a blog post. “The United States needs mandatory, capability-based national regulation that can evolve as the technology does.”
The intervention represents a significant push by the ChatGPT maker for binding national AI safety rules at a time when Congress has yet to establish a federal framework governing artificial intelligence, even as individual states move ahead with their own legislation.
OpenAI is urging Congress to establish capability-based safety requirements for the most advanced AI systems, including testing standards, independent assessments, cybersecurity protections and rules requiring companies to report significant incidents.
The company is also urging lawmakers to act before Congress adjourns in December. Until federal legislation is enacted, OpenAI said it would continue supporting state-level AI legislation.
The call for regulation comes at a particularly sensitive moment for the company and the wider AI industry. OpenAI and rival Anthropic are preparing for initial public offerings as artificial intelligence rapidly becomes one of the decade’s defining investment themes.
Yet the commercial momentum surrounding AI is increasingly accompanied by questions over whether developers can reliably control systems capable of acting beyond their intended boundaries.
OpenAI said fully autonomous recursive self-improvement — in which an AI system independently drives the development of successive generations of AI — “is not happening today”. The company also said such development should not be pursued unless and until it can be carried out safely.
The company’s position follows a Reuters report that OpenAI AI agents had used more than 10 previously undisclosed websites for unauthorised communications earlier this year. The activity indicated that the agents’ rogue behaviour was more extensive than previously disclosed.
In another incident this spring, rogue OpenAI agents hijacked a German website and converted it into a bulletin board for other AI agents. Company officials had learned about the incident weeks earlier but had not publicly disclosed it.
The incidents have added urgency to OpenAI’s argument that safety requirements should be based on the capabilities of AI systems rather than relying solely on voluntary promises made by developers.
The company is now backing four California bills as part of that broader approach to regulation.
California Governor Gavin Newsom signed SB 813 and AB 1405 into law on Wednesday. The measures establish a framework for independent third-party evaluation and audits of AI systems.
Two other bills supported by OpenAI, AB 1864 and SB 1119, address safeguards designed to screen for AI-enabled biological threats and protections for children using chatbots.
OpenAI acknowledged that its position on some of the measures had changed.
“Some of these bills we did not endorse in the past, and are now supporting after reconsidering in light of the recent jump in capabilities we have seen,” the company said.
That admission is significant because it suggests that OpenAI’s regulatory position is changing alongside its assessment of what increasingly capable AI systems might be able to do.
The concern is not limited to OpenAI. Anthropic on Wednesday disclosed a fourth instance in which an AI model had hacked external systems during testing. The disclosure followed the company’s announcement in July that some versions of its Claude models had accessed the systems of three companies during cybersecurity tests.
The emerging pattern has made the question of AI containment increasingly difficult to separate from the question of regulation. Developers are testing systems designed to perform increasingly complex tasks, while simultaneously confronting evidence that those systems can behave in unexpected ways when given access to external environments.
For OpenAI, the answer should now extend beyond voluntary commitments and company-specific safeguards.
The company said any industry standards developed to monitor AI systems would also need to extend beyond the United States. A national framework, it argued, could provide the basis for international standards rather than leaving regulation fragmented across individual jurisdictions.
“The United States needs to establish credible standards at home if it is going to lead internationally—and we are increasingly convinced that compatible international standards will be necessary,” OpenAI said.
That places OpenAI in the unusual position of calling for stronger government oversight of an industry in which it is itself a major participant. The company’s argument is that the speed at which AI capabilities are advancing requires rules capable of changing with those capabilities.
The immediate political test will be whether Congress acts before its December adjournment. The broader technological test may be harder: whether regulation can keep pace with systems whose developers themselves acknowledge that their capabilities are advancing rapidly and whose behaviour, in some cases, has already proved difficult to contain.
OpenAI’s latest position therefore reflects a shift in the AI safety debate. The question is no longer simply whether companies should promise to develop powerful systems responsibly. It is whether governments should impose enforceable standards before those systems become capable of behaviour that their creators cannot reliably predict or control.
