Fifteen Eighty Four

Revelations that swarms of AI agents, developed in OpenAI’s frontier AI lab, had broken out of their containment to coordinate an attack on data stored by Hugging Face, a machine learning solutions company, made the headlines in July 2026. Media coverage ranged from sober analyses of cybersecurity risks associated with frontier AI models to wild speculation about human extinction in the hands of machines. The so-called OpenAI-Hugging Face incident prompted Anthropic co-founder and CEO Dario Amodei to publish “We Must Pace the Frontier”, which basically argues for a slowdown in AI research until government regulation and/or business self-regulation could catch up with AI development – a call that drew support from other tech leaders, such as Sam Altman and Elon Musk, and elicited an angry reaction from US President Donald Trump.

Is Amodei right in demanding a slowdown? Based on my findings in Governing AI: A Primer, I believe that acknowledging the system-wide unintended consequences of emerging AI models is important and timely, as is Amodei’s advocacy for a combination of government regulation and business self-regulation. He is not the only, or even first, technology leader to seek tighter rules, but the context in which “We Must Pace the Frontier” was published has made its intended audience receptive to its message. Yet, if the current moment is to usher in a new era of safe AI development and use, the private and public sectors should adopt a principle-based, comprehensive approach to AI risks and harms, and work towards building a broad-based political alliance to turn principles to action.

It is indeed scary when AI agents find unauthorized paths in faraway corners of the web and pass messages among themselves to exploit cybersecurity vulnerabilities – though it should be noted that portrayals of this conduct as AI sentience are overblown. One could only imagine the catastrophic consequences if the frontier models in question were publicly available for customers to pursue their innocent – or not-so-innocent – goals with. While regulating frontier AI to contain downstream effects is important, it is not the only, or even the most pressing, risk arising from AI applications today. Chatbot hallucinations, disinformation campaigns, algorithmic amplification of psychological harm, mass surveillance, privacy concerns, and resource use stand out as harms that affect individuals and communities every day. Centering these real social, political, and economic problems in AI safety discourse is necessary to align AI with humanity’s collective well-being.

Amodei’s essay signals a commitment to doing the right thing if other leading American AI companies do the same – in his words, “to create a race to the top.” Governing AI: A Primer finds that businesses cannot get there through self-regulation alone. Consortia that bring together Big Tech companies conduct research and coordinate among members but lack the power to enforce ethical principles. Even the most celebrated examples of self-regulation, such as Meta’s Oversight Board that reviews content moderation decisions on Facebook and Instagram, are disallowed from questioning the harmful implications of that company’s core business model. What is worse is that Big Tech companies routinely disband teams and boards that are supposed to keep them ethically compliant. Thus, a race to the top is likely to fall victim to competition dynamics unless companies are obliged to comply with binding decisions enforced from outside – in a word, government.

In fact, Amodei’s essay reads like an invitation to other American tech leaders to support a new New Deal in which the federal government ensures rule-governed market competition to avoid a race to the bottom. One does not have to reinvent the wheel for sensible regulation: the AI Risk Management Framework developed by the National Institute of Standards and Technology in 2023, the Blueprint for an AI Bill of Rights drafted under the Biden administration, as well as non-US legislation, such as the European Union’s AI Act, provide a helpful albeit imperfect baseline. Alas, most tech leaders in the United States spent enormous energy and money during the Biden administration, the 2024 election campaign, and the early years of the second Trump presidency lobbying against legislative and administrative efforts to rein in AI harms. They cannot be trusted to lead AI governance.

Social movements tackling mass surveillance, data centers, disinformation, AI adoption in schools, and more, have been pushing for change and scoring victories at the local level. AI investors, tech leaders and influential politicians, that is, those who control money, know-how and political power, have tended to view those social movements as an obstacle to innovation. It is time to build a multi-stakeholder, cross-class coalition of tech workers, concerned citizens, academics, civic organizations, and yes, those tech leaders with a genuine interest in AI governance, to demand principled, comprehensive and binding rules to rule AI development and use.