97% of Americans believe artificial intelligence should be regulated, according to the latest Gallup survey. Yet the Trump administration had made AI deregulation one of its flagship promises, revoking in January 2025 all the safeguards put in place by Joe Biden. Three weeks later, it performs a 180-degree turn.
The trigger: Anthropic publishes its Mythos model on January 28, capable of identifying thousands of unknown cybersecurity vulnerabilities in critical American infrastructure in just minutes. Faced with this demonstration, even the most fervent opponents of regulation discover that certain AI capabilities transcend political divisions. Technological innovation forces ideology’s hand.
The Essentials
- 97% of Americans support AI regulation according to Gallup 2025, transcending all party divisions
- The Trump administration revoked all Biden AI decrees in January 2025 before reversing course after the Mythos release
- 1,208 AI bills were introduced in all 50 states in 2025, creating an unprecedented regulatory mosaic
- Anthropic Mythos detects cybersecurity vulnerabilities at a scale that exceeds human capacity for control
From the Promise of Deregulation to Forced U-Turn
The Trump administration had built its technology strategy on a total rejection of the “bureaucratic constraints” inherited from Biden. On January 20, inauguration day, Donald Trump signs an executive order repealing his predecessor’s Executive Order on AI Safety, eliminating with a single stroke the transparency obligations for models exceeding 10^26 FLOPS and safety tests before deployment.
Silicon Valley applauds. Shares of OpenAI, Anthropic, and tech giants climb. The message is clear: American innovation must no longer be bridled by institutional caution.
Except that on January 28, everything changes. Anthropic launches Mythos in limited beta for 500 selected users. Within 72 hours, the model identifies 47,000 vulnerabilities in federal systems, including 12,000 classified as “critical” by the Cybersecurity and Infrastructure Security Agency. Among them: flaws in Pentagon networks, undocumented backdoors in air traffic control systems, exploitable entry points in the electrical infrastructure of 23 states.
The problem doesn’t only lie in the volume of flaws discovered, but in the speed of their identification. Where federal cybersecurity teams would take months to audit these systems, Mythos delivers its analyses in real time. “We’ve created the perfect cyber warfare tool,” an Anthropic executive confides anonymously to the Washington Post. “If we can see these vulnerabilities, our adversaries can too.”
On January 31, Donald Trump urgently convenes the leaders of Anthropic, accompanied by the directors of the NSA and FBI. On February 2, the administration announces the creation of a temporary “AI Safety Board,” exactly the type of agency it had abolished three weeks earlier.
1,208 Bills and Impossible Federal Coordination
This federal reversal occurs in an already fragmented regulatory landscape. According to the count by MultiState, a legislative monitoring organization, 1,208 bills concerning artificial intelligence were introduced in all 50 American states in 2025, representing a 340% increase compared to 2024.
This explosion is explained by the void created by the repeal of Biden’s measures. Faced with temporary federal inaction, states took the lead. California maintains its AI Safety Act, Texas imposes mandatory audits for AI systems used in public services, New York requires algorithmic transparency for any AI processing personal data.
The result: an unmanageable legal patchwork for technology companies. A startup developing a voice recognition model must now navigate between 15 different regulations depending on the states where it operates. Global antitrust catches up with tech giants, but this time the fragmentation comes from within the American system itself.
“We’re witnessing the Balkanization of the American AI ecosystem,” analyzes Sarah Mitchell, director of the Special Competitive Studies Project. “Each state creates its own standards, its own control agencies, its own definitions of what constitutes a ‘dangerous’ AI. It’s exactly the opposite of what the administration wanted by eliminating federal regulation.”
This fragmentation benefits international competitors. While American states argue over their respective regulations, China deploys a coordinated national approach, the European Union finalizes its AI Act, and Singapore attracts AI companies with a unified and predictable regulatory framework.
Public Opinion Forces Elected Officials’ Hands
The overwhelming support of 97% of Americans for AI regulation reveals a rare consensus in a polarized political landscape. This proportion even exceeds approval for popular measures like universal healthcare (89%) or climate change action (72%).
More remarkable still: this support crosses party lines. 94% of Republican voters approve AI regulation, compared to 98% of Democrats and 97% of independents. “This is probably the most consensual political outcome we’ve measured since the September 11 attacks,” notes Jeffrey Jones, editorial director of Gallup.
This convergence is explained by concrete and immediate concerns. 78% of respondents fear that AI will replace their jobs, 71% worry about information manipulation, 69% dread mass surveillance. These fears transcend traditional ideological divisions between regulation and free market.
The Mythos affair crystallizes these anxieties. By revealing thousands of critical vulnerabilities, Anthropic involuntarily demonstrates that AI can become a weapon of mass destruction against national infrastructure. “We’ve created a monster,” headlines the Wall Street Journal. “And now we have to learn to control it.”
This popular pressure partly explains the speed of Trump’s reversal. According to internal administration polling revealed by Politico, 82% of Republican voters believe the president “must act quickly” to regulate the most powerful AIs. Even his electoral base, otherwise committed to deregulatory theories, demands government intervention in the face of cybersecurity risks.
Anthropic Involuntarily Becomes De Facto Regulator
In developing Mythos, Anthropic thought it was demonstrating the defensive capabilities of advanced AI. The company wanted to show how its models can identify and correct security flaws faster than any human team. The reverse result: it demonstrates that AI can become the most sophisticated cyber warfare tool ever created.
“We’ve broken the internet,” Dario Amodei, CEO of Anthropic, soberly summarizes during a Congressional hearing on February 5. “Mythos identifies vulnerabilities we didn’t even know to look for. If this technology falls into the wrong hands, it can paralyze any country in a few hours.”
Mythos’s capacity indeed exceeds anything that existed before. The best human hackers find on average 2 to 3 critical vulnerabilities per month in a complex system. Mythos identifies several thousand per day, with a precision rate of 94%. More troubling: the model develops intrusion techniques that cybersecurity experts had never considered.
Faced with this power, Anthropic finds itself de facto regulating its own technology. The company limits access to Mythos to 50 hand-picked organizations: government agencies, large cybersecurity companies, selected academic institutions. Each use is subject to real-time audit, each discovered vulnerability must be immediately reported to the competent authorities.
“We’re doing the job the state should be doing,” acknowledges Amanda Askell, head of AI safety at Anthropic. “But we have neither the democratic legitimacy nor the regulatory expertise to make these decisions. This is a responsibility that far exceeds a private company.”
This forced self-regulation reveals the inadequacy of existing frameworks. Laws on the export of sensitive technologies don’t cover AI models. Regulations on cyber weapons remain vague regarding vulnerability discovery tools. The junior pyramid wavers in major American law firms facing automation, but here it’s the entire cybersecurity sector that is revolutionized in a few weeks.
The Search for Balance Between Innovation and National Security
The new AI Safety Board created urgently by Trump illustrates the difficulty in reconciling technological innovation and national security. Composed of 12 members from the private sector, administration, and academic world, this agency must define a framework for “dual-use” AIs within 90 days.
First task: classify models according to their potential for harm. Mythos falls into the “cyber weapon AI” category, requiring prior authorization from the Department of Homeland Security before any deployment. Models capable of generating biological agents or controlling critical infrastructure fall under the same classification.
This risk-based approach is inspired by the British model, where the AI Safety Institute evaluates each advanced system before market launch. But the American application remains more permissive: only AIs presenting “catastrophic risks” are subject to prior control.
“We’re seeking the balance point,” explains Chris Krebs, former CISA director appointed to the Safety Board. “Too much regulation kills American innovation. Too little regulation kills America outright. Mythos shows us that this red line really exists.”
The technology industry remains divided. AI startups applaud a lighter framework than Biden’s measures, giants like OpenAI and Google fear their Chinese competitors will pull ahead while America regulates. Meta pleads for industry self-regulation, Microsoft calls for coordinated international standards.
This tension reflects AI and Solow’s paradox, an example of the American economy: how to maximize AI’s productivity gains without suffering its negative externalities? The answer will determine whether the United States retains its technological edge or loses it to more coordinated approaches elsewhere.
Toward a New Bipartisan Technology Doctrine
Trump’s reversal on AI could herald a lasting bipartisan consensus on regulating advanced technologies. For the first time in years, Republicans and Democrats converge on the need for federal oversight, even if they diverge on its modalities.
This convergence goes beyond AI alone. Recent bipartisan agreements on semiconductor research, quantum research funding, and fusion energy investments show an America capable of consensus facing major technological challenges.
“Trump discovered what Biden already knew: you can’t let the market regulate alone technologies that can destroy society,” analyzes Marietje Schaake, former European Parliament member specializing in digital issues. “This bipartisan awareness opens the way to a more mature American technology policy.”
The question remains open: will this emerging regulation be fast enough to govern innovation, or will AI companies continue to set the pace? With 1,208 bills underway in the states and a federal Safety Board under construction, America is experimenting in real time with a new form of technology governance. The Mythos accident will have at least had this merit: forcing debate on questions that no one can ignore anymore.
Sources
- Americans Prioritize Safety, Data Security on AI - Gallup
- MultiState - Analysis of AI bills by state in 2025
- Special Competitive Studies Project - Report on American regulatory fragmentation
- Washington Post - Interviews with Anthropic leaders (February 2025)
- Politico - Internal Trump administration polls on AI