ATOMIC Act Would Force OpenAI, Google, Anthropic to Submit to Nuclear-Risk AI Testing

Utah Rep. Celeste Maloy introduced the AI Threat Output and Monitoring Incident Containment Act (ATOMIC Act) on August 6, 2026, alongside Rep. Sara Jacobs of California. The bipartisan bill would require researchers at Department of Energy national labs to test the largest AI models for nuclear-related security risks, including jailbreaking and dangerous information generation, before those models are publicly deployed. Companies in scope include OpenAI, Google DeepMind, Anthropic, xAI, and Meta.

What happened

Detail	Fact
Bill name	AI Threat Output and Monitoring Incident Containment Act (ATOMIC Act)
Introduced	August 6, 2026
Sponsors	Rep. Celeste Maloy (R-UT) and Rep. Sara Jacobs (D-CA)
“Advanced AI” compute threshold	More than 10^26 floating-point operations used in training
Investment threshold for compliance	$2 billion in AI development over the previous five years
Testing location	Department of Energy national laboratories
Named companies in scope	OpenAI, Google DeepMind, Anthropic, xAI, Meta

The ATOMIC Act would create a formal team at DOE national labs to evaluate frontier AI models before public release. Researchers would conduct what Maloy’s military legislative assistant Andrew Caprio described as “red teaming”: probing for flaws, jailbreaking vulnerabilities, and any capabilities the model has to assist in nuclear weapons development or the compromise of nuclear facilities.

The bill’s definition of an “AI nuclear incident” covers a wide range of scenarios, from a model generating step-by-step weapons instructions to a foreign adversary gaining unauthorized access to or manipulating an AI system. Scheming behavior related to nuclear stockpiles is also listed.

Why it matters

This is one of the first serious legislative attempts in the US to tie AI safety testing directly to nuclear security infrastructure. Rather than creating a new agency, the bill routes oversight through DOE national labs, institutions that already hold the highest security clearances and deep nuclear expertise.

The compute-based threshold is significant. At 10^26 floating-point operations, the bar is set at the very top of current model training runs, which means only a handful of organizations would be directly regulated today. Smaller AI companies are indirectly covered in theory because, as Caprio noted, many are fine-tuned on outputs from the larger frontier models. Testing the big five, the argument goes, effectively stress-tests their derivatives.

The bill also promises to protect trade secrets, which matters for companies that worry the red-teaming process could expose proprietary model architectures or training data. That protection is likely what makes the “relationship with large developers” framing possible, as Caprio put it.

For businesses that rely on AI tools, including AI integration for internal workflows, this kind of structured government evaluation signals that frontier models will face increasing external scrutiny. That scrutiny may slow some model releases at the margin, but it also builds the public and regulatory trust that sustains long-term adoption.

What is red teaming in this context?

Red teaming, as used here, means researchers actively try to make an AI model do something it should not do. In the nuclear context that means probing whether a model can be manipulated into providing restricted data, generating weapons instructions, or supporting a foreign actor seeking access to nuclear material. The goal is to find those failure modes in a controlled lab setting, then build mitigation strategies before the model goes public.

This practice already happens informally. Anthropic, OpenAI, and Google DeepMind all run internal safety evaluations. The ATOMIC Act would add a mandatory, independent layer conducted by government researchers with security clearances that private red teams do not hold. You can follow ongoing coverage of AI safety legislation at Lumien’s AI news desk.

Our take

The ATOMIC Act is narrow and technically grounded, which puts it ahead of most AI regulation attempts. Anchoring scope to a specific compute threshold avoids the endless “what counts as AI” debate that has derailed other bills. Routing testing through national labs rather than a new bureaucracy is also smart: the expertise already exists there.

The weakness is the same one Caprio partially acknowledges. If a future model achieves dangerous capability at lower compute costs, as DeepSeek demonstrated is possible with different training approaches, the threshold becomes a loophole. The bill’s framers seem aware of this and treat the threshold as a starting point, not a permanent line.

For AI product builders and agencies, the practical message is: frontier model releases may carry longer pre-deployment review windows if this passes. That is worth building into timelines for any project that depends on cutting-edge model features. It is not a reason to panic, but it is worth watching.

Source: Bing News · Anthropic