Anthropic's most powerful AI "suddenly died" overnight: Insider Story of Amazon's Betrayal and White House's 90-Minute Ultimatum
According to Dynamic Beating monitoring, Claude Fable 5 was just released for three days when it was urgently taken offline by the White House due to a fatal disclosure from the major shareholder Amazon. On Thursday, Amazon's CEO Andy Jassy called Treasury Secretary Bessemer directly, claiming that his research team had successfully used trigger words to bypass Fable 5's outer defenses, reach the underlying Mythos 5 model, induce it to generate output that could exploit software vulnerabilities, and automatically create testing scripts, ultimately transforming them into exploit validation (PoC) code through a multi-step process.
Friday morning Eastern Time, the vulnerability incident directly alarmed the top levels of the White House. White House Chief of Staff Susie Wells urgently convened a closed-door meeting to discuss response plans, while Treasury Secretary Bessemer, en route to Houston, remotely dialed in during his trip. During the meeting, the White House submitted jailbreak test data to the National Security Agency (NSA) for evaluation, and the NSA later confirmed to the White House that the evidence was conclusive, and the vulnerability had substantial network weapon-grade lethality. After the meeting, the White House urgently sought out Anthropic's founder and CEO Dario Amodei, only to find that he was "unreachable." Some White House officials claimed that Amodei was attending a "silent meditation retreat" at the time, but Anthropic later angrily refuted this claim as "pure fabrication," explaining that Amodei had quickly connected within an hour of receiving the contact request at noon, and the company had proactively provided other executives as substitutes during his absence.
The crisis erupted completely at 1:00 p.m. Eastern Time on Friday. The White House spared no effort, with up to half of the core Cabinet members, including Secretary of Commerce Lutnick, launching as many as three rounds of intense phone pressure on Amodei. In multiple calls, National Security Advisor Kaincross demanded strongly that Amodei voluntarily take down the model. Amodei vehemently defended himself, insisting that this was only a bypass for a specific scenario, not a catastrophic "universal jailbreak," and accusing the government of overreacting. This self-defense struck White House officials as extremely ironic. Amodei had previously likened AI risks to an "atomic bomb" in policy lobbying, but now, faced with a real vulnerability, he was trying to downplay it. In the confrontation, Bessemer directly warned Amodei that he was making a "very stupid decision" and forcibly imposed a 90-minute physical takedown deadline.
The two sides' arguments fell into a deadlock afterward: White House officials claimed that the export control was a reluctant move made "after several hours of pleading yielded no results," while sources close to Anthropic denounced the White House as brutal, stating, "They directly threw a 90-minute physical death deadline without providing any technical details, leaving no room for negotiation." After the negotiations broke down, the Trump administration formally issued an emergency export control order at 5:30 p.m. on Friday, and the capitulating Anthropic was eventually completely physically disconnected at 10:00 p.m. that night. Subsequently, the company issued an official statement, condemning the government's actions as "unfair, opaque, and disrespectful of technical facts."
As Amazon, Anthropic's largest external investor, deliberately shorts its AI ally in which it holds a significant stake, the underlying motive is a complete reversal driven by vested interests. Despite previously investing $13 billion in Anthropic, Amazon this year poured a whopping $50 billion into Anthropic's competitor, OpenAI, and was granted permission to resell its latest model on AWS. This massive transaction directly transformed the once deadly denunciation of Anthropic into Amazon's showcase of sweeping obstacles aside for its new favorite, forcefully escorting its $50 billion gamble.
For the Trump administration, this whistleblowing comes as a godsend. The White House has long held grievances against this liberal-leaning AI startup: Anthropic heavily hires officials from the Biden era, and its founder has publicly criticized Trump. A deeper conflict erupted back in March when Anthropic staunchly refused to allow its models to be used for large-scale domestic surveillance and lethal autonomous weapons development, leading the Pentagon to label it a supply chain risk, resulting in ongoing lawsuits. Therefore, under the persuasion of cybersecurity officials, President Trump ultimately signed the executive order.
This export ban subsequently produced an extremely absurd side effect: due to restrictions on any "foreign national" access, even a large number of core R&D personnel from allied countries such as the UK and Canada within Anthropic were cut off from system privileges, causing an immediate halt to its technical team. For Anthropic, which is actively preparing for a fall IPO, this heavy blow from the White House almost prematurely spelled doom for its business plan, while the vacant market space is swiftly being cannibalized by OpenAI from behind.
