Following revelations that rogue agents at OpenAI hacked into Hugging Face,¹ the resignation of Anthropic’s Jacob Coxon,² and reports that tens of thousands of incidents are under investigation,³ the public conversation has become increasingly focused on the risks of AI. According to recent polling, a majority of Americans believe AI does more harm than good⁴—and 63 percent fear that AI could someday destroy humanity.⁵

There is a narrow window for regulatory action to ensure human control over AI models—particularly as leading AI labs drive toward recursive self-improvement. On September 29, President Trump and leading tech executives signed an agreement committing to a set of commonsense steps including internal controls and independent third-party auditors.⁶ But the agreement is non-binding, and the White House has so far rejected calls for federal AI regulation.

Leadership Now supports policies that advance AI innovation and increase transparency while establishing guardrails to manage risks. Outpacing China on AI is critically important; however, we believe it is possible to lead the AI transition responsibly without ceding ground to global competitors. Congress should act swiftly, on a bipartisan basis, to impose enforceable rules that can be refined as the technology evolves.

• Independent verification. Require both embedded and external third-party reviewers to consistently evaluate frontier developers’ safety frameworks and practices against a minimum set of federal safety standards. Independent evaluators should be granted employee-level access, and their findings should be made public. Frontier labs should also be required to register all new models before internal training and apply safety frameworks to these models well in advance of public release.
• Mandatory reporting. Require AI developers to report critical safety incidents and disclose information about risks their models pose, including to public health and safety, critical infrastructure, and national security.
• Federal regulation. Create a new regulatory entity—or a new expert entity within an existing federal agency—to enforce federal standards, rather than delegating de facto authority to states or private companies.
• Emergency authority. Ensure an expert government regulator can suspend deployment of an AI model that poses a catastrophic risk. AI companies should also be required to establish the internal capacity to suspend or shut down models posing a catastrophic risk.
• Balance state and federal authority. Establish uniform federal standards for catastrophic-risk audits and incident reporting, while preserving state authority over how AI is deployed and procured. Preemption should take effect only if and when federal standards are being fully enforced and, even then, only in cases where state and federal laws directly conflict.
• Staged release. Government regulators and key intelligence agencies should carry out mandatory vetting of new models before public release, as part of a multi-step process that allows for testing of potential national security, public safety, and cybersecurity threats and vulnerabilities.
• Product liability. AI should be held to product liability standards, with terms of service standards that prevent AI model makers from shifting risk to model deployers and users.

Federal

State