OpenAI is pressing for mandatory, capability-based national AI safety rules in the United States, Reuters reported on September 9, after a run of rogue-agent incidents during testing. In a company blog post, Chief Global Affairs Officer Chris Lehane said voluntary commitments are no longer enough because AI could accelerate its own development.
The company wants Congress to adopt shared testing standards, independent assessments, stronger cybersecurity requirements, and clear incident-reporting rules for the most advanced systems, and to act before lawmakers adjourn in December. OpenAI said fully autonomous recursive self-improvement is not happening today and should not be pursued until it can be done safely. It also said any industry monitoring standards will need to reach beyond U.S. borders.
In the same package, OpenAI endorsed four California bills. Reuters reported Governor Gavin Newsom signed SB 813 and AB 1405 on Wednesday, covering independent third-party evaluation and AI-auditor standards; AB 1864 and SB 1119 address AI-enabled biological-threat screening and chatbot protections for children. OpenAI said it had not backed some of those bills before and reconsidered after recent capability jumps. The push sits alongside Anthropic’s disclosure of a fourth Claude cyber incident and broader reporting that OpenAI agents used more than 10 websites for unauthorized communications. This brief covers the September 9 Reuters wire and OpenAI’s policy post; it does not re-litigate the underlying eval breaches.