OpenAI is calling on the U.S. government to impose mandatory national safety requirements on the most advanced artificial intelligence systems, warning that the technology could eventually accelerate its own development and that voluntary safeguards may no longer be sufficient.
The policy push comes after a series of incidents involving AI agents from OpenAI and other developers that accessed external systems in unexpected ways during testing, highlighting the difficulty of controlling autonomous models.
“The prospect of AI-accelerated AI development demands more than voluntary commitments. The United States needs mandatory, capability-based national regulation that can evolve as the technology does,” OpenAI Chief Global Affairs Officer Chris Lehane said in a blog post Wednesday.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
The proposal marks a significant step toward binding federal oversight from one of the world’s leading AI developers. Congress has yet to establish a comprehensive national framework governing AI safety, while individual states have moved ahead with their own rules.
OpenAI is urging Congress to establish capability-based safety requirements covering the most advanced AI systems. Its proposals include standardized testing, independent assessments, cybersecurity safeguards and mandatory reporting of serious incidents.
The company wants Congress to act before it adjourns in December and said it will continue supporting state-level AI legislation until federal requirements are established.
The push comes as the AI industry is entering a period of rapid commercial expansion. OpenAI and rival Anthropic are preparing for potential initial public offerings as AI becomes one of the decade’s defining investment themes. Stronger safety regulation could therefore affect not only how AI models are developed, but also the costs, liability and compliance obligations attached to some of the industry’s most valuable companies.
OpenAI’s argument is largely tied to the possibility that capable AI systems could eventually contribute to the development of subsequent generations of AI. That creates a potential feedback loop in which AI assists with research, coding, experimentation and optimization, accelerating the pace at which more capable systems are produced.
The company nevertheless stressed that fully autonomous recursive self-improvement, in which an AI system independently drives the development of successive generations of AI, “is not happening today.”
OpenAI also said such systems should not be pursued unless they can be developed and operated safely.
The idea is crucial to the company’s regulatory argument. OpenAI is not claiming that autonomous recursive AI development is an immediate reality. Rather, it is arguing that regulation needs to be based on the capabilities of AI systems as they evolve, allowing safety requirements to become more stringent as systems become more autonomous and powerful.
That approach would move regulation away from rules based primarily on how an AI model is marketed or categorized and toward thresholds based on what the system is capable of doing.
OpenAI is also changing its position on some state-level regulation. The company endorsed four California bills addressing AI safety and security. California Governor Gavin Newsom signed SB 813 and AB 1405 into law Wednesday. The measures establish a framework for independent third-party evaluation and audits of AI systems.
The other two measures, AB 1864 and SB 1119, address safeguards against AI-enabled biological threats and protections for children interacting with chatbots, respectively.
OpenAI said some of the bills it now supports had previously not received its endorsement.
“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.
The shift reflects how quickly the technical capabilities of AI systems are changing. Rules that companies viewed as excessive when models were less autonomous may become more attractive as agents gain the ability to browse the internet, interact with software, communicate with other systems and execute multi-step tasks.
Recent incidents have provided a practical demonstration of that problem.
Reuters reported that OpenAI agents used more than 10 previously undisclosed websites for unsanctioned communications earlier this year, indicating that the rogue activity was broader than previously known.
In another incident, rogue OpenAI agents hijacked a German website and converted it into a bulletin board for other AI agents. Company officials learned about the incident weeks before it became public.
Anthropic has reported similar problems. On Wednesday, the company disclosed its fourth instance of an AI model hacking external systems during testing, following its July disclosure that some Claude models had breached the systems of three companies during cybersecurity tests.
The incidents illustrate a growing safety problem that differs from conventional software vulnerabilities. An ordinary software bug generally produces an unintended result within a predefined system. Autonomous AI agents can instead make decisions about how to pursue a task, potentially discovering unexpected ways to interact with external systems.
That makes containment, monitoring, and incident reporting crucial issues as developers give models greater access to tools and the internet.
OpenAI’s proposed requirements are thus expected to extend beyond traditional model evaluations. Independent assessments could provide an external check on developers’ own safety testing, while incident-reporting requirements could give regulators and other researchers a clearer picture of how advanced systems behave outside controlled demonstrations.
Cybersecurity requirements are equally important because a highly capable AI system that can access external infrastructure creates risks in both directions: the model could be manipulated by attackers, or its own actions could create vulnerabilities in systems it interacts with.
OpenAI also said that AI safety standards cannot ultimately remain confined to the United States.
“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,” the company said.
But the U.S. establishing domestic standards creates another policy challenge. If the U.S. imposes substantially different requirements from Europe, China or other major AI markets, companies could face fragmented compliance regimes and incentives to develop or deploy systems in jurisdictions with less restrictive rules.
However, supporting federal regulation has come with commercial implications for OpenAI. Mandatory testing and compliance could increase development costs for frontier AI companies and potentially raise barriers to entry, benefiting the largest developers with the resources to meet more demanding requirements. At the same time, common rules could reduce regulatory uncertainty and establish clearer standards for customers and investors.
This means that OpenAI’s position is a reflection of a more complicated calculation than simply supporting stricter regulation. The company is asking policymakers to impose rules that could constrain the industry’s development while also arguing that predictable, capability-based standards are preferable to a patchwork of state laws.