White House and Congress weigh AI security rules as Pentagon and firms sign classified deals

What This is

• The Trump administration has used export controls and national-security claims to restrict access to Anthropic’s newest frontier AI models while Congress and the White House debate broader AI security rules, and the Pentagon has disclosed classified AI use
• The Commerce Department directed Anthropic to bar foreign nationals from using Fable 5 and Mythos 5, and Anthropic then disabled access more broadly because it said it could not reliably screen users by nationality.
• U.S. officials cited a reported method to bypass model safeguards and expose powerful cybersecurity capabilities, with reporting that Amazon flagged the issue and that the government gave Anthropic little time to comply.
• Anthropic disputes the severity and scope of the alleged vulnerability, saying it found only narrow or minor issues and warning the directive is overbroad and could affect future model releases and business plans.
• More than 100 cybersecurity executives and experts urged the administration to ease or lift the restrictions and asked for a transparent, scientific process for assessing AI security risks.
• The White House and Anthropic have been in talks about creating standardized benchmarks for assessing the severity of security breaches in advanced AI models, involving Commerce and national cyber officials.
• The episode has pushed lawmakers to seek briefings and revive efforts for federal AI rules, including ideas such as mandatory reviews of advanced models and other guardrails.
• Pentagon testimony filed in court says xAI’s Grok Gov model is used inside Project Maven and helped support U.S. strikes on Iran, including deploying over 2,000 munitions to 2,000 targets within 96 hours.
• The Grok disclosure appears in a Justice Department filing defending an xAI data center against an NAACP Clean Air Act lawsuit over gas turbines and alleged pollution in majority-Black Memphis neighborhoods, with the DOJ arguing the facility supports
• The wider AI boom is driving massive compute buildouts, including Meta contracting for about 1.6 gigawatts of AI computing capacity from data-center firm Crusoe across two U.S. sites.

Why This Matters

• Applying export-style controls to AI software could set a durable precedent for how governments restrict who can use powerful models, shifting key decisions from companies and courts to executive agencies.
• The dispute spotlights a core governance problem: security claims often rely on technical tests and classified context that are hard to audit publicly, raising legitimacy and due-process concerns.
• Military adoption of advanced AI for targeting increases the urgency of oversight because mistakes, biased inputs, or rushed automation can translate directly into lethal outcomes and escalation risks.