Jailbroken AI Sparks National Security Crackdown

Circuit board with glowing brain chip design

The Trump administration just proved it is willing to pull the plug on Big Tech’s most powerful AI when national security is on the line.

Story Snapshot

Commerce ordered Anthropic to cut off its newest AI models to all foreign users after a jailbreak raised cyberwar fears.
The directive used tough export-control powers normally aimed at weapons and advanced chips, signaling AI is now treated as strategic tech.
Anthropic says the exploit was “minor,” but Amazon and other testers reportedly unlocked deeper hacking capabilities inside Fable 5.
The move exposes a wider fight over who controls frontier AI: elected government or unelected tech executives.

Trump Team Treats Frontier AI Like a Strategic Weapon

When Anthropic rolled out its new Fable 5 and Mythos 5 models on June 9, tech insiders called them the most powerful public AI systems yet, especially for cybersecurity work and complex coding. Within days, Commerce Secretary Howard Lutnick sent a letter ordering Anthropic to halt exports of those models to all foreign destinations and to foreign nationals everywhere, citing fears they could aid Chinese or Russian military intelligence.^[1]^[2] That order instantly turned a commercial product into a national security asset in the eyes of Washington.

The Trump administration used authority under the 2018 Export Control Reform Act and the broader export-control framework that was tightened in 2024–2025 to cover advanced AI model “weights.”^[1] Those rules already let the Bureau of Industry and Security require licenses for sending top-tier, closed-weight models anywhere in the world, not just to adversary countries. By invoking that same toolbox against Anthropic’s remote-access models, Commerce signaled that even cloud AI outputs will be treated like sensitive technology when cyberwar risks are involved.

Jailbreak Scare: How a Safety Model Became a Cyber Weapon Risk

According to officials who spoke with reporters, concerns shifted from theory to reality once Amazon’s AI security team and several other companies testing Fable 5 found a “jailbreak” that bypassed its cyber safeguards.^[2] Fable 5 was supposed to route risky cybersecurity, biology, or chemistry queries to a weaker model, Opus 4.8, using extra safety classifiers. But testers reportedly discovered a prompt method that re-opened Mythos-level cyber abilities, letting the model identify software vulnerabilities that could threaten core internet infrastructure.^[2]^[5]

In its public statement, Anthropic downplayed the exploit as narrow and low-impact, saying the demonstration only surfaced a small number of already-known, simple bugs that other public models can also find without any jailbreak.^[2] The company added that outside red teams and bug bounties had never found a “universal” jailbreak that stripped away all safeguards. Still, Commerce told Anthropic it viewed the exploit as serious enough that misuse by foreign intelligence agencies could not be ruled out, especially if the same trick were repeated at scale.^[1]^[2]^[5]

Why a Foreign-National Ban Took the Models Offline for Everyone

The Bureau of Industry and Security did not order Anthropic to shut down the models for Americans. Instead, it banned use of Fable 5 and Mythos 5 by any foreign national, whether inside the United States or overseas, unless Anthropic obtained a license.^[1]^[2] Because Anthropic’s application programming interfaces cannot reliably check citizenship in real time across the globe, the company responded by disabling both models for all users, worldwide.^[4] That single technical constraint turned a targeted foreign-access ban into a de facto global shutdown.

Legal analysts note this is the first time Commerce has used export authority this way on a live, widely deployed AI service rather than on chips or model weights moving across borders.^[1]^[4] Some export-control experts question whether remote access to a model counts as an “export” under traditional rules, since the software itself may never leave U.S. servers.^[1] But years of expanding controls on advanced chips, model weights, and even model outputs laid the groundwork for this step, and the Trump team is clearly willing to test those boundaries in court if needed.

Anthropic Fights Back While Trump Tightens the Rules of the Game

This clash did not come out of nowhere. Commentators point out that the Pentagon and Anthropic have been locked in a months-long dispute over the company’s refusal to allow its systems to be used for mass domestic surveillance and fully autonomous weapons.^[3]^[4] A separate national security memorandum by President Trump already told defense agencies to end contracts with AI vendors that “limit” government use in critical missions, a move widely seen as aimed at Anthropic.^[3] In that context, the export order looks like both a security step and a pressure tactic in a larger power struggle.

At the same time, Trump signed an executive order on June 2 to build a formal process for vetting top AI models for up to 30 days before release, focused on cyber risks and critical infrastructure protection.^[1]^[2]^[5]^[7] The order directs national security agencies to create a classified benchmarking system for “covered frontier models” and an AI cybersecurity clearinghouse to find and patch software vulnerabilities faster.^[1]^[5]^[7] Importantly for conservatives who dislike heavy-handed regulation, the order states it does not create a mandatory licensing regime for all new AI models.^[1]^[7]

What It Means for Conservative Readers and the Road Ahead

For Americans worried about hostile regimes, the Anthropic case shows the administration sees advanced AI the same way it sees hypersonic missiles or top-end chips: too important to hand to China or Russia on a silver platter.^[1] Existing “catch‑all” export controls already bar U.S. persons from helping foreign weapons programs with any U.S.-origin technology when they know it will be used for missiles or weapons of mass destruction. Frontier AI models that can discover zero‑day software flaws now fall into that same strategic conversation, not just tech hype.

Still, this episode also warns about the blunt tools Washington is using. A narrow cyber exploit, whose real-world danger is debated, ended up knocking out useful AI tools for law‑abiding users worldwide because the legal and technical frameworks were not ready. That is why several legal and security scholars argue for clearer, risk‑based AI laws that focus on truly sensitive domains—like cyber operations and advanced weapons design—without turning every powerful model into a permanent emergency. The Trump administration’s next steps will determine whether this clampdown stays targeted, or becomes a template future bureaucrats might use in ways that threaten innovation and free expression.