Three days after its launch, Anthropic's flagship Claude Fable 5 was yanked offline by the White House — sparked by a fatal tip from its largest shareholder, Amazon. On Thursday, Amazon CEO Andy Jassy directly called Treasury Secretary Bessent, claiming his research team had used prompt injections to bypass Fable 5's outer guardrails, reach the underlying Mythos 5 model, and coax it into generating output capable of exploiting software vulnerabilities — eventually turning it into a working proof-of-concept exploit.
By Friday morning, the breach had reached the highest levels of the White House. Chief of Staff Susie Wills convened an emergency closed-door meeting, with Bessent dialing in from his car en route to Houston. The White House sent the jailbreak data to the NSA for evaluation; the agency confirmed the evidence was solid and the vulnerability had genuine cyber-weapon-level destructive power. After the meeting, aides frantically tried to reach Anthropic founder and CEO Dario Amodei — only to find he was "unreachable." One White House official claimed Amodei was at a "silent meditation retreat," a charge Anthropic angrily dismissed as "pure fabrication." The company insists Amodei called in within an hour of being contacted at noon, and that other executives had been offered as stand-ins during his brief absence.
The crisis came to a head at 1:00 PM Eastern. Half of the core cabinet — including Commerce Secretary Lutnick — launched three rounds of intense phone pressure on Amodei. On the conference line, National Security Adviser Cairncross demanded that Anthropic voluntarily take the model down. Amodei pushed back hard, arguing it was a narrow bypass, not a catastrophic "general jailbreak," and accused the government of overreacting. The White House found his defense bitterly ironic: Amodei had repeatedly likened AI risk to "the atomic bomb" in policy lobbying, yet now tried to minimize a real vulnerability. During the exchange, Bessent warned Amodei he was making "an extremely stupid decision" and imposed a 90-minute physical takedown deadline.
The two sides' accounts diverged sharply afterward: White House officials claimed the export control was a last resort after "hours of pleading" fell on deaf ears; sources close to Anthropic accused the administration of bullying — "they just threw a 90-minute physical death warrant with no technical details, leaving zero room for negotiation." With talks broken, the Trump administration signed an emergency export control order at 5:30 PM Friday. Anthropic reluctantly complied, fully disconnecting the model by 10:00 PM. The company then issued a statement blasting the move as "unfair, opaque, and disrespectful of technical reality."
Amazon — Anthropic's biggest external investor — had effectively shorted its own heavily funded AI ally, a wholesale defection driven by capital interests. Though Amazon had previously poured $13 billion into Anthropic, this year it invested a staggering $50 billion into Anthropic's rival OpenAI and won the right to resell OpenAI's latest models on AWS. That massive deal turned Amazon's lethal whistleblowing into a loyalty badge for its new darling — clearing the path for a $50 billion gamble.
For the Trump administration, the tip landed perfectly. The White House had long resented the liberal-leaning AI startup: Anthropic had hired numerous Biden-era officials, and its founders had openly criticized Trump. The deeper conflict erupted back in March, when Anthropic refused to let its models be used for mass domestic surveillance or lethal autonomous weapons — earning it a Pentagon designation as a supply-chain risk, with lawsuits ongoing. Cybersecurity advisors persuaded Trump to sign the ban.
The order also had an absurd side effect: because it restricted access by "any foreign national," even key Anthropic researchers from allied countries like the UK and Canada were locked out, paralyzing the technical team. For a company racing toward an IPO this fall, the White House's blow effectively killed its business plan — and the resulting market vacuum is being rapidly filled by the rival it left behind: OpenAI.