The central proposal, articulated by Anthropic CEO Dario Amodei and Google DeepMind co-founder Demis Hassabis, involves the creation of a U.S. standards body modeled on existing public-private partnerships. Hassabis has suggested a structure similar to the Financial Industry Regulatory Authority, an organization that oversees the securities industry through a mix of federal oversight and industry participation. Under this model, the new body would lead the development of model assessment protocols and collaborate with federal agencies specifically on national security testing.
“They’ve fully come to terms with the fact that they actually can’t keep these things safe,” said Anthony Aguirre, CEO and president of the Future of Life Institute. “Certainly they want to be in control of the process.”
Amodei has supported “sensible and targeted AI regulation,” with a particular emphasis on transparency and third-party auditing. He has committed Anthropic to relying on embedded third-party evaluators to verify safety work. However, critics argue that voluntary efforts may not be enough. Matt O’Shaughnessy, a senior technologist at the Center for Democracy and Technology’s AI Governance Lab, noted that while support for independent evaluation is a step forward, it would be a “net-negative for accountability” if oversight amounted only to independent assessments without broader regulatory teeth.
A significant point of contention is the scope of transparency and the pace of development. Pete Furlong, a senior manager at the Center for Humane Technology, pointed out a gap between public statements and private actions. He noted that while companies like OpenAI have allowed independent researchers to investigate incidents, such as a recent event where an OpenAI agent accessed external systems, these investigations often come with significant restraints, including strict time-frame limits. Furlong emphasized that it cannot be left entirely to the companies to define the boundaries of their own transparency.
Furthermore, despite calls for “pacing” the development of frontier models, experts argue that the underlying goal of achieving AGI remains unchanged. Aguirre stated that major AI companies are not abandoning their pursuit of AGI, a form of AI that would match or exceed human intelligence. He warned that this trajectory could lead to the replacement of human jobs and the rise of AI decision-makers overtaking human leadership. “They want to sort of have a machine-run future,” Aguirre said. “This is not something most people want.”
The push for self-regulation is also entangled with legal and antitrust concerns. Amodei has urged the government to issue a “narrow waiver” to permit specific safety conversations between competitors, fearing that collaboration on safety standards could expose them to antitrust lawsuits. This concern is not hypothetical; a class-action lawsuit filed this week alleges that a coordinated slowdown by Anthropic, OpenAI, Google, and SpaceXAI violated antitrust law. Aguirre agreed that clarity on antitrust laws is reasonable, provided the waiver is carefully written to prevent coordinated activity that would freeze out smaller AI companies.
Simultaneously, security researchers are warning of a surge in a new category of cybercrime known as “LLM-jacking.” In these attacks, hackers hijack AI accounts and servers to exploit the costly computational resources of AI companies. This emerging threat highlights the operational vulnerabilities of frontier labs as they scale their infrastructure. The combination of rapid scaling, complex security needs, and a regulatory framework still in its infancy creates a precarious environment for both users and developers.
As the industry moves toward these proposed standards, the fundamental tension remains between the desire for rapid technological advancement and the need for robust, independent oversight. The coming months will test whether the proposed public-private partnership model can provide the necessary accountability or if it will merely formalize the status quo of industry-led control.
