How Amazon CEO Triggered the Fable 5 Shutdown

The week of June 9 started with Anthropic celebrating one of its biggest launches ever. It ended with the company pulling its two most capable models from every user on Earth. New reporting from Fortune reveals the chain of events that produced that outcome - and the central role played by an Amazon executive, a phone call, and a cybersecurity access program that had grown faster than anyone was tracking.

TL;DR

Amazon CEO Andy Jassy flagged the Fable 5 jailbreak to White House officials on June 11 during a call scheduled for a completely different topic
Commerce Secretary Lutnick gave Anthropic a 90-minute ultimatum: fix the vulnerability or pull the models
SK Telecom's appearance on Project Glasswing's access list - added without prior government approval - eroded trust with national security officials
Dario Amodei rejected both options Sacks offered; talks had reached an impasse by June 16
Security researchers say the White House's condition for redeployment is technically unachievable

The Call That Started It

On June 11, Amazon CEO Andy Jassy was on a pre-scheduled call with White House officials. The agenda covered something unrelated to AI models. Before the call ended, Jassy mentioned that Amazon researchers had documented a method for bypassing Fable 5's safety guardrails - the model Anthropic had released two days earlier.

White House officials encouraged Jassy to raise the issue directly with Treasury Secretary Scott Bessent. He did so the same day.

By the following evening, the matter had moved from a briefing at Treasury to the Commerce Department. Commerce Secretary Howard Lutnick signed a directive to Anthropic at 5:21 PM on June 12. Anthropic had 90 minutes to respond.

A Competitor Raising the Alarm

The escalation path matters because it wasn't the only one available. Amazon and Anthropic are closely linked - Amazon has a $4 billion investment in Anthropic and serves as the company's primary cloud infrastructure partner. Jassy chose to raise the vulnerability with White House officials rather than flag it to Anthropic directly. That decision set a faster and more official clock ticking than any internal conversation would have.

Amazon didn't comment on Jassy's role in the events. Anthropic declined to discuss the internal timeline.

What Project Glasswing Had to Do With It

The jailbreak alone might not have produced a global shutdown. What escalated the response was a second finding: the expanded roster of Anthropic's Project Glasswing.

Project Glasswing is Anthropic's controlled-access program giving a select group of organizations access to Claude Mythos - a model with strong cybersecurity reasoning capabilities positioned for national security applications. The program had been built on close coordination with government stakeholders.

By early June, Anthropic had expanded the access list to roughly 50 new organizations without obtaining prior approval from US national security officials. One of those additions was SK Telecom, South Korea's largest wireless carrier. SK Telecom denied having any ties to China, but its appearance on the Glasswing roster without prior clearance raised questions in Washington about who else had received access to Mythos capabilities.

Why the List Expansion Changed the Calculation

When the jailbreak report arrived, it landed on top of a preexisting concern. If Anthropic had been extending Glasswing access without consultation, the scope of exposure was unclear. The vulnerability report and the access list expansion arrived together, and together they produced a different response than either would have triggered alone.

The revelation of the ballooning Glasswing roster and the inclusion of the Korean telecom firm damaged trust between the US national security establishment and Anthropic.

The Fortune investigation, published June 18, describes this combination - jailbreak plus unauthorized expansion - as the key factor in the government's decision to move as fast as it did.

Anthropic's public statement about the Fable 5 and Mythos 5 suspension Anthropic published a detailed public statement defending its position on the jailbreak standard, arguing that applying it industrywide would halt all new frontier model releases. Source: anthropic.com

The Ultimatum and Amodei's Decision

Lutnick's directive gave Anthropic two choices: fix the jailbreak, or take Fable 5 and Mythos 5 offline. The 90-minute window made real-time remediation impossible. Anthropic pulled both models globally before 10 PM on June 12.

CEO Dario Amodei rejected both options on principle. The jailbreak in question involved asking the model to review a codebase and fix software flaws - a legitimate, common use case. Anthropic's public statement described the finding as a "narrow, non-universal jailbreak" that didn't meet the bar for a full commercial recall.

"If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers," Anthropic said.

David Sacks's Signal

White House AI adviser David Sacks offered a more optimistic read shortly after the shutdown, saying the administration viewed the issue as "easily resolved" and that the matter was now in Anthropic's hands.

By June 16, four days later, talks had stalled. The export controls that shut down access for European and other foreign users remained in place. Sacks has been serving in an advisory capacity since stepping down from his formal White House role earlier this year; his public signals carried visibility but not authority to close a deal.

As of June 19, Anthropic's managing director of international, Chris Ciauri, told reporters that the models would be available "in the coming days" - the most specific positive signal since the June 12 order, but still without a firm date.

The Zero-Jailbreak Problem

The practical question is what Anthropic actually has to do to get Fable 5 and Mythos 5 back online.

Security researchers contacted by multiple outlets gave the same assessment: a jailbreak-free frontier model doesn't exist. "You cannot guarantee that a model will remain jailbreak-proof forever," Martin Riley, CTO at cybersecurity firm Bridewell, told Gadget Review. "Anyone promising that is selling something."

Fortune's June 18 investigation was the first to reveal Andy Jassy's role and the unauthorized Glasswing expansion Fortune's June 18 deep-dive was the first to connect Andy Jassy's White House call, the Glasswing access expansion, and the resulting government response into a single account. Source: fortune.com

Anthropic's public argument rests on exactly this point. The company isn't disputing that the jailbreak exists. It's disputing the standard being applied - that a single narrow vulnerability in a commercial model justifies a global recall. That standard, Anthropic argues, can't be met by any lab, and applying it uniformly would mean no new frontier model ever ships.

The original shutdown covered in our June 13 report described what happened. The Fortune investigation explains why it happened the way it did. What it doesn't explain is how it ends. The government needs a condition Anthropic can actually meet. Anthropic needs one it can accept without establishing that a routine jailbreak finding gives regulators the power to take its products offline on 90 minutes' notice. Neither side has offered that yet.

Sources: