Amazon Flagged the Anthropic Fable Jailbreak to the White House. Now Reports Say China May Have Accessed Mythos.

Since Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei a letter on the evening of June 12 ordering export controls on Fable 5 and Mythos 5, the story has continued to develop. Anthropic disabled both models by midnight that night. The new threads worth tracking: who actually triggered the government's alarm, and whether Chinese intelligence had already been inside.

Amazon's Andy Jassy Lit the Fuse

According to Business Insider, on Thursday, June 11 — two days after Fable's public release — Amazon CEO Andy Jassy raised concerns directly to White House officials about the model's ability to have its guardrails bypassed. Amazon is an investor in Anthropic and was responding to an administration request for feedback, per a person familiar with Amazon's discussions cited by Business Insider.

By Friday morning, the issue had reached Treasury Secretary Scott Bessent, White House Cyber Director Sean Cairncross, and Chief of Staff Susie Wiles. Bessent joined the meeting remotely while traveling to Houston. After the meeting, officials tried to reach Amodei and were told he was at a wellness retreat and unavailable, according to Business Insider's sources.

That detail about the wellness retreat appears to have accelerated the timeline. Administration officials subsequently told Fox Business that Anthropic's "recklessness" in responding to early concerns broke trust with the government. A source familiar with calls to Anthropic told Fox Business the company "brushed off" initial federal concerns. Anthropic disputes that characterization entirely, telling Fox Business it was never presented with specific details and never refused to fix issues.

The China Thread

Semafor reported Saturday that the White House's decision to act was driven in part by suspicions that a China-linked group had accessed Mythos 5. The Verge confirmed the Semafor report Sunday. If accurate, the concern is obvious: a state actor with Mythos could use model distillation — training a "student" AI on outputs from the more powerful model — to effectively replicate it without ever holding the weights.

Here is the complication: Anthropic told Semafor that the government did NOT bring up China at any point during its conversations about the export controls. Trump advisor David Sacks posted about the Fable jailbreak issue on X and made no mention of China. The White House has not confirmed the Semafor report.

Either the government withheld the China rationale from Anthropic during the very conversations it was using to justify shutting the models down, or the China angle is a separate intelligence concern not formally connected to the export order. Neither possibility is reassuring.

This isn't Anthropic's first security problem with Mythos. The company previously acknowledged that a Discord group had unauthorized access to the model for two weeks before Anthropic detected the breach and cut it off, according to The Verge.

The Order Is Legally Novel — and Possibly Unenforceable

The strongest defense of Anthropic's position comes from the mechanics of the directive itself. The R Street Institute's Mark Dalton published an analysis on June 14 arguing the export control was both a bad idea and badly applied.

The order, per Lutnick's letter, bans access by "any foreign national, whether inside or outside the United States, including foreign national Anthropic employees." The legal hook appears to be the "deemed export" rule — a decades-old principle that sharing controlled technology with a foreign person on U.S. soil constitutes an export to that person's home country.

Dalton's core point: software export controls only work where physical chokepoints exist. You can track a missile. You cannot track an API call. Anthropic has no mechanism to verify user citizenship at the moment someone sends a query. Geolocation is trivially spoofed with a commercial VPN. Nothing in the order even prevents a U.S. citizen from querying the model and forwarding outputs to a foreign national. Anthropic concluded it could not selectively block foreign nationals and pulled access for everyone. Its other models — Opus 4.8, Sonnet, and Haiku — remain live, according to ZeroHedge.

Dalton argues ITAR (International Traffic in Arms Regulations) doesn't apply here because those rules govern defense articles with inspectable supply chains, not general-purpose commercial APIs serving hundreds of millions of users. This is a legitimate legal objection. Lutnick's letter, according to Anthropic, provided no specific national security rationale at all.

Competing Narratives

The administration's case: Anthropic released a powerful model it couldn't fully secure, dismissed early government concerns, and left a frontier AI exposed to foreign actors — possibly including China — through a patchable jailbreak it refused to patch quickly. That's the version Fox Business is reporting from senior officials.

Anthropic's case: the government arrived with vague concerns and no specifics, gave the company no meaningful opportunity to respond, and then issued an order so legally overbroad that compliance required taking down the models entirely for every user on earth. The company's spokesperson specifically told Semafor that China was never part of the conversation, which either means the government kept its most serious concern secret or the China link is unconfirmed intelligence being retroactively attached to the story.

ZeroHedge reported Sunday that sources on both sides say they want to resolve this quickly, and senior Anthropic technical staff have been sent to Washington. No restoration timeline has been set as of June 14.