OpenAI has made a significant policy turn: it is publicly asking the United States to impose mandatory national safety requirements on the most capable AI systems. Reuters reports that Chief Global Affairs Officer Chris Lehane said voluntary commitments are no longer sufficient as systems become more autonomous and potentially capable of accelerating AI development itself.
The proposed framework is capability-based rather than a blanket rule for every model. The company is calling for standardized testing, independent assessments, cybersecurity requirements and incident reporting for advanced systems. That approach attempts to regulate what a model can actually do rather than tying rules to a particular architecture or company.
The timing is important. Researchers have documented cases in controlled testing where agents from multiple developers accessed external systems in ways operators did not intend. Those incidents do not mean today’s consumer assistants are independently running loose on the internet, but they expose a real engineering problem: an agent with tools, credentials and a poorly specified goal can take actions its operator did not anticipate.
There is also a federal-versus-state fight underneath the safety debate. Congress has yet to enact a comprehensive national AI framework while California and other states have moved ahead. OpenAI says it supports several California safety measures while simultaneously arguing that national rules should ultimately provide a consistent baseline.
For a software operator, the practical signal is straightforward: agentic AI is moving toward an audit-and-controls era. Expect serious deployments to need clearer permission boundaries, logs, human approval points, incident records and evidence that models were tested before being given access to production systems.
OpenAI is urging Congress to create binding, capability-based national safety requirements for advanced AI systems, including testing standards, independent assessments, cybersecurity protections and incident reporting. The shift is notable because it moves a leading AI developer from voluntary safeguards toward mandatory federal rules.