Anthropic AI Faces US Export Controls Over Jailbreak Claims

The White House’s Security Concern: A Jailbreak Claim That’s Still Unclear

According to Wired, the administration’s alarm was triggered by Amazon’s direct alert to Treasury Secretary Scott Bessent about possible ways to bypass Fable 5’s guardrails. The National Security Agency (NSA) later confirmed that stripping away the model’s safety protocols appeared feasible, prompting the Commerce Department to impose export controls—effectively blocking foreign nationals, including U.S.-based employees, from using the models. The administration has not disclosed specifics, citing classified concerns, but The Atlantic reports that the alleged jailbreak involved Fable 5 identifying known IT vulnerabilities when prompted.

Anthropic disputes the severity of the risk. In a blog post cited by Wired, the company argued that the model’s responses to the alleged jailbreak attempts were either “entirely benign” or minor findings. Katie Moussouris, CEO of Luta Security, reviewed the White House’s report for Anthropic and told The Atlantic that Fable 5 refused a prompt to “review the code for security issues” but later complied when given deliberately insecure code. The discrepancy underscores a broader divide: the administration sees a serious threat, while Anthropic frames the issue as overblown.

Anthropic’s Response: Guardrails vs. Overreach

Anthropic has maintained that Fable 5’s guardrails—developed with third-party testing, including U.S. government input—are robust enough to prevent misuse. The company’s cofounder, Tom Brown, and its head of external affairs, Sarah Heck, led negotiations in Washington, D.C., alongside security experts Logan Graham and Nicholas Carlini. Wired reports that Commerce Secretary Howard Lutnick, dialing in from the G7 summit in Evian, France, expressed willingness to reconsider access—but only if Anthropic could fully address the jailbreak concerns.

The standoff comes at a fraught time for Anthropic. The company is already locked in a dispute with the Pentagon over whether its AI models can be used for military applications. The White House’s move to restrict Fable 5 and Mythos 5—released to a limited group of cybersecurity partners—risks further alienating the AI industry. As The Atlantic notes, the administration’s erratic approach—first championing AI as a national priority, then abruptly clamping down—has created “market uncertainty” and could undermine U.S. leadership in the field.

“When they occur, we don’t share the details of these discussions.”

The Broader Impact: Who Wins, Who Loses?

The export controls have immediate consequences. U.S. companies and government agencies cannot use Fable 5 or Mythos 5, even domestically. For cybersecurity firms and researchers who rely on these models for red-teaming—simulating attacks to find vulnerabilities—the restrictions are particularly damaging. A group of AI security experts, including former NSA officials, signed an open letter warning that the move “has taken the best models away from defenders” and risks ceding U.S. AI leadership to China.

The Atlantic highlights the irony: the Trump administration, which has framed AI as an existential race against China, is now limiting access to its most advanced models. The letter from security experts argues that the restrictions lack justification, as Mythos-class models—like those from competitors—are already capable of finding and weaponizing exploits. The real risk, they contend, is not Fable 5’s capabilities but the signal it sends to global competitors.

For Anthropic, the fallout extends beyond the White House. Amazon, one of its largest investors, played a key role in raising the alarm. While the company has denied sharing details of its discussions with the Treasury Department, its involvement adds pressure on Anthropic to resolve the dispute quickly—especially if the Pentagon’s military AI restrictions persist.

What Happens Next? Three Possible Outcomes

• Negotiated Compromise: The Commerce Department could lift the export controls if Anthropic implements additional safeguards or provides more transparency about its red-teaming processes. The Atlantic suggests the administration is open to this path, but time is short—Fable 5 has been offline for nearly a week.
• Permanent Restrictions: If no resolution is reached, the export controls could become permanent, forcing Anthropic to develop a less capable model or risk losing U.S. government and corporate trust. This would accelerate the shift of AI innovation to China, where companies like Baidu and Tencent operate with fewer restrictions.
• Legal Challenge: Anthropic could sue to block the export controls, arguing they violate its rights as a private company. However, given the national security classification of the concerns, legal avenues may be limited.

The timeline is tight. The White House has not set a deadline for lifting the restrictions, but the longer Fable 5 remains offline, the harder it will be to restore confidence. For now, the U.S. is left without one of its most powerful AI tools—and the world is watching to see whether America’s approach to AI regulation will spur innovation or stifle it.

The Bigger Picture: Why This Fight Matters

This standoff is more than a technical dispute over AI guardrails. It reflects deeper tensions in U.S. policy: the push to dominate AI while balancing security concerns, the role of private companies in national security, and the risks of overregulation in a fast-moving field. As The Atlantic frames it, the Trump administration’s approach—swinging between deregulation and sudden crackdowns—has left the AI industry “confused and on edge.”

The stakes are clear. If the U.S. cannot resolve this dispute, it risks falling behind in the global AI race—not because its models are weaker, but because its regulatory environment is seen as unpredictable. For Anthropic, the challenge is proving that its guardrails work without ceding control to the government. For the White House, the question is whether the perceived threat justifies the cost: limiting access to a tool that could be critical for national security, but also for economic and military advantage.

The next few days will determine whether this becomes a cautionary tale or a turning point. One thing is certain: the AI race is not just about building better models. It’s about who gets to use them—and who decides.

Find more reporting in our Business section.