Trump Reverses Course on AI Safety Testing

President Donald Trump's administration has quietly begun reinstating AI safety testing requirements it previously dismantled — a dramatic reversal that sources say was triggered by classified demonstrations of advanced AI capabilities, including a system reportedly codenamed Mythos, that alarmed senior officials. The move effectively validates the Biden-era approach to AI governance that Trump spent months publicly ridiculing.

The reversal marks one of the most significant policy U-turns of the current administration and signals that even the most deregulation-minded political leaders cannot ignore the accelerating pace of AI advancement.

Key Takeaways

• Trump officials are reinstating portions of Biden's AI safety executive order (EO 14110) that they revoked in January 2025
• A classified AI capability demonstration reportedly 'spooked' senior administration officials into action
• The new framework will require pre-deployment safety testing for frontier AI models above certain compute thresholds
• Industry leaders from OpenAI, Anthropic, Google DeepMind, and Meta were consulted in recent weeks
• The policy shift represents a tacit acknowledgment that Biden's approach to AI safety had merit
• Implementation timelines remain unclear, but initial guidance could arrive within 60 days

What Is Mythos and Why Did It Alarm the White House?

Mythos refers to what multiple reports describe as an advanced AI capability demonstration that was presented to senior Trump administration officials during a national security briefing. While specific details remain classified, sources familiar with the matter indicate the demonstration showcased AI systems capable of autonomous planning, sophisticated deception, and self-directed goal pursuit at levels that exceeded what policymakers had anticipated.

The demonstration reportedly left several officials visibly unsettled. One source described the reaction as a 'come-to-Jesus moment' for an administration that had built its tech policy around removing regulatory barriers.

What makes the Mythos incident particularly significant is its timing. It arrived at a moment when frontier AI labs are racing to develop AGI-level systems, with OpenAI, Anthropic, and Google DeepMind all claiming they are within striking distance of artificial general intelligence. The gap between what Washington understood about AI capabilities and what the labs were actually building had grown dangerously wide — precisely the scenario Biden's original executive order was designed to prevent.

Biden's AI Safety Framework: What Trump Originally Scrapped

In October 2023, President Joe Biden signed Executive Order 14110, the most comprehensive AI governance directive ever issued by the U.S. government. The order required companies developing frontier AI models to share safety test results with the federal government before deployment. It established reporting thresholds based on computational resources, mandated red-teaming exercises, and created a framework for ongoing monitoring.

Trump revoked the order on his first day back in office, calling it 'innovation-killing bureaucracy' that would hand America's AI lead to China. The decision was cheered by portions of Silicon Valley and criticized by safety researchers who warned it left the U.S. without guardrails during the most critical period of AI development.

Key elements of Biden's original framework included:

• Mandatory safety reporting for models trained above 10^26 FLOP
• Red-team testing requirements before public deployment
• Government authority to review dual-use foundation models for national security risks
• Establishment of the AI Safety Institute (AISI) within NIST
• Requirements for watermarking AI-generated content

The irony now is unmistakable. Many of these provisions are reportedly being reconsidered — in some cases nearly word-for-word — by the same administration that dismantled them.

The Industry Response: Relief Mixed With Frustration

Reaction from the AI industry has been complex. Several major companies, including Anthropic and Google DeepMind, had continued voluntary safety testing even after Trump removed the federal mandate. For these firms, the policy reversal represents validation of their approach but also frustration over months of regulatory whiplash.

Dario Amodei, CEO of Anthropic, has consistently argued that safety testing and competitive advantage are not mutually exclusive. His company's Responsible Scaling Policy — which gates deployment of increasingly powerful models behind safety evaluations — has served as an informal industry standard even in the absence of federal requirements.

OpenAI's position is more nuanced. The company initially welcomed Trump's deregulatory stance but has recently signaled openness to structured safety requirements. CEO Sam Altman reportedly participated in recent White House consultations, advocating for a framework that is 'rigorous but not bureaucratic.'

Meta, which has pursued an open-source strategy under Mark Zuckerberg, faces perhaps the most complex calculus. Pre-deployment safety testing for open-weight models raises unique challenges, since once a model is released, control over its use becomes effectively impossible.

Why This Reversal Matters Beyond Washington

The implications of Trump's U-turn extend far beyond American borders. The European Union's AI Act, which took effect in stages throughout 2024 and 2025, established the world's most comprehensive AI regulatory framework. The U.S. decision to abandon safety testing had created a transatlantic regulatory gap that complicated international cooperation on AI governance.

With the U.S. now moving back toward mandatory safety testing, several important dynamics shift:

• International alignment on AI safety standards becomes more feasible
• The G7 Hiroshima AI Process gains renewed momentum with U.S. re-engagement
• China's AI governance approach faces greater pressure to converge with Western standards
• Companies operating globally face a more consistent regulatory landscape
• The UK AI Safety Institute and its U.S. counterpart can resume meaningful collaboration

For developers and businesses building on top of frontier models, this shift signals that safety compliance is not optional — regardless of which party holds power. Companies that maintained robust testing protocols during the regulatory gap are now positioned advantageously compared to those that cut corners.

What This Means for Developers and Businesses

Practical implications are already emerging for teams building AI-powered products. Companies deploying frontier models should anticipate new compliance requirements within the next 2 to 3 months. Those who maintained safety documentation and red-teaming practices will face minimal disruption.

Startups and smaller developers using API access from major providers like OpenAI or Anthropic are largely insulated, since the testing requirements apply primarily to foundation model developers rather than downstream users. However, companies fine-tuning open-weight models for sensitive applications — healthcare, finance, legal — should prepare for enhanced scrutiny.

The cost of safety testing is non-trivial. Industry estimates suggest comprehensive red-teaming and evaluation for a frontier model costs between $5 million and $20 million, depending on scope. This represents a meaningful barrier for smaller labs but a rounding error for companies like Google and Microsoft that spend billions on model training.

Looking Ahead: The New AI Safety Landscape

The coming months will determine whether Trump's reversal represents genuine commitment to AI safety or a temporary reaction to a frightening demonstration. Several key milestones to watch include:

First, the administration must decide whether to formally reinstate portions of Biden's executive order or craft entirely new directives. The former would be politically embarrassing but operationally efficient. The latter allows for face-saving but risks delays during a critical period.

Second, the role of the AI Safety Institute remains uncertain. Under Biden, AISI was positioned as the federal government's primary technical authority on AI evaluation. Trump initially sidelined the organization, and its future staffing and mandate will signal how seriously the administration takes implementation.

Third, congressional action looms. Bipartisan AI safety legislation, including updated versions of proposals from Senators Chuck Schumer and Todd Young, could codify testing requirements into law — making them resistant to future executive reversals.

The Mythos incident, whatever its precise details, appears to have accomplished what years of expert testimony, academic papers, and industry warnings could not: convincing a skeptical administration that the risks of ungoverned AI development are real, immediate, and potentially irreversible. Whether that conviction translates into durable policy remains the defining question for American AI governance in 2025.