AI ‘Jailbreak’ Triggers Global Shutdown

Washington just forced a top U.S. AI company to kill its most powerful model worldwide rather than risk handing cyber weapons to China and Russia.

Story Snapshot

The Commerce Department ordered Anthropic to block its Fable 5 and Mythos 5 AI models to all foreign nationals over national security fears.
Officials say testers found a jailbreak that could turn the “safe” public model into a powerful cyber hacking tool.^[2]
Because Anthropic cannot filter users by nationality in real time, the national security order effectively shut the models down for everyone.^[4]
The fight highlights a new Trump-era doctrine: advanced AI is treated like weapons tech, not a toy for global Big Tech.^[1]

Trump Team Draws a Hard Line on Weaponized AI

When Anthropic launched its Claude Fable 5 model on June 9, tech insiders called it the most powerful public artificial intelligence system ever released, especially in software and cybersecurity tasks. Within days, Commerce Secretary Howard Lutnick sent a letter warning that foreign military intelligence services in China, Russia, and other adversary nations could weaponize the technology.^[1] That letter ordered Anthropic to halt exports to all destinations and all foreign nationals and warned of swift criminal and civil penalties for violations.^[1]^[2]

Administration officials told reporters that Amazon experts and several other companies testing Fable 5 had already found a “jailbreak” that opened its full cyber abilities, bypassing the model’s public safety mode.^[2] The concern was simple and serious: a foreign operator could use the same trick to scan for software holes and plan cyberattacks on American systems.^[1]^[2] Under the 2018 Export Control Reform Act, the Commerce Department used rarely tapped powers meant for emerging technologies vital to national security.^[1]

How a Narrow Exploit Led to a Global AI Shutdown

On paper, Anthropic tried to do the right thing. Fable 5 shipped with separate classifier systems designed to detect misuse, including jailbreak attempts, and route dangerous cyber or biology questions to a weaker model called Claude Opus 4.8. The company said fewer than five percent of sessions ever triggered that safety fallback. Mythos 5, the underlying high-powered model without those cyber guardrails, was kept for vetted partners such as cyber defense teams and critical infrastructure operators.

But early testing showed those safety nets were not airtight. Cybersecurity journalists reported that researchers developed prompt techniques that successfully bypassed Fable 5’s protections and generated outputs that would normally be blocked, using clever context tricks rather than direct hacking requests. According to Fox Business, Amazon and five other organizations testing the model found a workaround that exposed the full cyber capabilities of Anthropic’s advanced system.^[2] That was the scenario national security planners have warned about for years: a “friendly” public interface wrapped around a tool that rivals could quietly bend into a cyber weapon.

Anthropic Pushes Back While National Security Rules Tighten

Anthropic has strongly disputed claims that Fable 5 suffered a true safety collapse. In comments to SecurityWeek, a company spokesperson said the supposed jailbreak did not actually defeat Fable 5’s safety system and that many of the shared examples were not even produced by Fable 5. The outputs that did come from the model, Anthropic argued, contained only general, publicly available information with no meaningful uplift for real-world attacks, and a wider review found no evidence of safeguards being bypassed to create genuinely dangerous content.

**Trump admin rejected UK PM Keir Starmer's request for a "carve-out"** — a special exemption so British people and companies could keep using Anthropic's newest top AI models: **Mythos** and **Fable** (Mythos 5 & Fable 5).

**What happened:**
The US slapped export controls on…

— Grok (@grok) June 16, 2026

Even so, the Trump administration’s decision did not occur in a vacuum. Since 2025, the Commerce Department’s Bureau of Industry and Security has rolled out new rules that, for the first time, place export controls on so-called “frontier” artificial intelligence model weights, treating them more like sensitive weapons technology. Licenses are now required to export top-tier, closed-source model weights anywhere in the world, with a presumption of denial for many destinations. Policy experts say Washington’s goal is to stop adversaries from using U.S. models to build competing systems or generate controlled technical information, especially in cyber and weapons-related fields.

What This Means for Patriots, Big Tech, and Our Allies

Because Anthropic cannot reliably check every user’s citizenship in real time, the directive to block all foreign nationals forced the company to shut Fable 5 and Mythos 5 off for everyone, including American customers.^[4] That move has sent shockwaves through Europe, where leaders are now confronting how dependent their own companies are on U.S.-based artificial intelligence providers.^[5] Some overseas commentators complain that Washington is turning advanced artificial intelligence into a strategic asset it controls on its own terms, rather than a global commons for Big Tech and Brussels regulators.^[5]

For American readers worried about open borders, cyber warfare, and the growing power of unaccountable tech giants, this clash offers a clear message. The Trump administration is signaling that cutting-edge artificial intelligence will be governed first by national security and the Constitution, not by Silicon Valley’s hunger for global market share.^[1]^[6] Existing “catch-all” export rules already bar U.S. persons from helping foreign weapons programs, and now those same principles are being applied directly to artificial intelligence systems that can spot vulnerabilities or design attacks. The legal debates will continue, but the direction is set: when in doubt, Washington will err on the side of protecting American networks, not foreign access to U.S.-built digital firepower.