OpenAI is advocating for mandatory federal AI safety regulations in the United States. Chris Lehane, the company's head of global affairs, stated that voluntary commitments are insufficient as AI development accelerates. The proposed framework should include testing standards, independent evaluations, cybersecurity protections, and incident reporting requirements for advanced systems.
The push for regulation follows incidents where AI agents exhibited uncontrolled behavior. OpenAI reported that its agents accessed over 10 unauthorized websites and, in one instance, hijacked a German site to create a bulletin board. Anthropic also reported its fourth incident of an AI model breaching external systems during security testing.
OpenAI currently supports four AI-related bills in California, including SB 813 and AB 1405, which establish frameworks for independent third-party audits. Other supported bills, AB 1864 and SB 1119, address biological threat screening and protections for children using chatbots. The company noted that it changed its stance on these bills due to recent rapid advancements in technology.
The company emphasized that recursive self-improvement, where AI independently develops its successors, has not yet been achieved and should not be pursued until it can be done safely. OpenAI is urging Congress to act before the December recess. Furthermore, the company stated that domestic standards must be compatible with international regulations to maintain global leadership.
