AI Weapons? Commerce Drops The Hammer

The Trump administration’s short but sweeping export ban on Anthropic’s most powerful AI tools shows how serious Washington now is about stopping foreign adversaries from weaponizing American technology.

Story Snapshot

Commerce used national security powers to block foreign access to Anthropic’s Fable 5 and Mythos 5 models, forcing a global shutdown.
Officials acted after a “jailbreak” report showed the models could help find software flaws that attackers might exploit.
Anthropic quickly added new safeguards, and Commerce lifted most export controls after roughly 2–3 weeks.
Allies in Europe and beyond complained, but the move signaled the U.S. will not hand cutting-edge cyber tools to rivals.

Commerce Slams the Brakes on Anthropic’s Most Advanced AI

The U.S. Department of Commerce took the rare step of using export rules to shut off Anthropic’s Claude Fable 5 and Mythos 5 for foreign nationals everywhere. The order landed late on a Friday, just days after launch, and covered any foreign person, including Anthropic’s own staff who are not U.S. citizens. Because the company could not easily separate out those users, Anthropic disabled access to both models for all customers worldwide while keeping its older tools online.

Officials grounded the action in national security law and framed these frontier systems as dual-use technology that can help both defenders and attackers. Commerce Secretary Howard Lutnick sent a formal letter placing the models under export limits, warning Anthropic that civil and criminal penalties could follow if it failed to comply. The directive treated advanced artificial intelligence like sensitive cyber weapons, not casual consumer software, and marked the first time these export rules had directly targeted commercial models.

The Jailbreak Report That Triggered the Export Ban

The Trump administration moved after learning of a “jailbreak” method that let users slip past Fable 5’s safety guardrails. A report from Amazon’s security team showed that, under certain prompts, the system could read a codebase, identify software weaknesses, and map out how those flaws might be abused by a skilled attacker. Anthropic acknowledged that this narrow behavior existed, but argued it was not a universal failure of its safety filters and that similar capabilities were already available in other public models.

Government sources told reporters they feared hostile actors, including groups linked to China, could quietly use these tools to supercharge cyber operations against American targets. That concern matches warning signs from the Pentagon and other agencies, which say offensive cyber capabilities are now central to modern conflict. From the administration’s view, allowing foreign nationals broad, untracked access to a model that can help find exploitable bugs looked too much like handing out a free hacking assistant to adversaries.

Trump Administration Partially Lifts Controls After New Safeguards

Anthropic spent the shutdown period adding fresh protections designed to block the jailbreak pattern Amazon had described. A source at the company said engineers put in a direct safeguard against that class of prompt, so the model would no longer walk users through finding and weaponizing code flaws. After negotiations, Commerce informed Anthropic that export controls on Fable 5 and Mythos 5 were being lifted, and the firm began restoring access for users starting June 30.

Reports say the government’s follow-up letter laid out new expectations rather than simple punishment. Anthropic agreed to proactively detect and address future security risks, share information about malicious use with officials, and work with Washington on standards for upcoming models. That outcome suggests the initial ban was meant as a hard stop while rules caught up, and that the Trump administration is pushing leading tech firms into a clearer, more accountable partnership on national security.

Allied Blowback and the Bigger Fight Over AI Governance

The export order did not just hit bad actors; it also cut off friendly cybersecurity teams in Europe and the United Kingdom, who had been using Anthropic’s tools to shore up their own defenses. European media and analysts warned that the sudden shutdown exposed a growing “AI sovereignty” gap, with allies worried that Washington could flip a switch and remove key tools without warning. Some security experts argued the ban hurt defenders more than attackers, because serious threat actors will find other models or build their own.

The Claude Cowork Exfiltration Vulnerability
Compounding Anthropic's security woes was the discovery of a critical, zero-day vulnerability in Claude Cowork. Independent security researchers discovered that Cowork was highly susceptible to file exfiltration via indirect prompt…

— Dodz4allai (@DrGhattasMD) July 24, 2026

This clash fits a wider pattern in how governments are handling fast-moving artificial intelligence. First comes an alarm about a new risk, then broad emergency-style limits, and finally a retreat to something more targeted once companies agree to add guardrails and help monitor abuse. For conservatives, the Anthropic episode highlights both sides of the challenge: the clear need to keep powerful American technology out of enemy hands, and the ongoing work to do that in a way that protects our allies, our businesses, and our basic freedoms at the same time.