New reporting from Axios, Fortune, and Semafor has filled in what happened behind the scenes in the hours before Anthropic’s Fable 5 was pulled from service on June 12 — and the picture is more complicated than the initial export control directive suggested.

The trigger was Amazon. Amazon researchers called White House administration officials on Thursday evening (June 12) with a live demonstration showing how they had prompted Anthropic’s Mythos-class model to generate restricted information about cyberattacks. The call happened hours before the Commerce Department issued its formal export control directive. By the time Anthropic received a letter from the government, the company had 90 minutes to comply.

Anthropic pulled both Fable 5 and Mythos 5 offline. As of June 16, both models remain unavailable.

The Amazon Factor

Amazon’s role in this story is complicated. The company is Anthropic’s largest cloud partner — AWS committed $25 billion to Anthropic in 2025, one of the largest technology infrastructure partnerships in the industry. Amazon Bedrock serves Anthropic’s models to enterprise customers at scale. Commercially, Amazon and Anthropic are deeply tied together.

And yet it was Amazon that brought the jailbreak demonstration to the White House.

The act itself exposes a reality most enterprise AI buyers have not fully factored into their risk models: your AI vendor’s cloud partners, infrastructure providers, and investors may have interests that diverge from your vendor’s — and those interests can affect your access.

Whether Amazon’s move was driven purely by national security concern or other motivations, the outcome is the same. A third-party company demonstrated a jailbreak on another company’s model and brought that demonstration to the government. That sequence caused immediate, unplanned disruption for every enterprise running workflows on Fable 5.

Why the CEO Said No

According to reporting by Tom’s Hardware, Anthropic was informed about the jailbreak finding and had an opportunity to address it before export controls were applied. CEO Dario Amodei declined, characterising the jailbreak as not a serious risk. Trump AI adviser David Sacks publicly confirmed this account, framing Anthropic’s refusal to act as the reason the government stepped in.

Anthropic’s own public statement told a different story: the vulnerability was narrow, previously known, and comparable to issues that exist across other AI models currently in production. The company called the government’s action “a misunderstanding” and said it was working to restore access.

Both things are consistent: a vulnerability can be technically minor and still prompt decisive government action when the policy context is right. The technical argument is not over. But while it unfolds, businesses that depended on those models have no access.

The Chinese Access Layer

Semafor and BusinessToday reporting indicates that fears about Chinese access to Mythos were a significant underlying factor in the White House’s decision to move quickly and broadly. The export control directive was written to prevent any foreign national from accessing the model — a scope so wide that Anthropic disabled both models entirely rather than attempt to filter by nationality in real time.

Anthropic has pushed back on this framing, with a spokesperson saying the White House did not specifically raise Chinese access in the conversations around the Fable 5 jailbreak. The two accounts are not necessarily contradictory: the Amazon demonstration may have been the proximate cause, while concerns about Chinese access provided the policy motivation to act decisively rather than negotiate.

What This Means for Business

The Fable 5 shutdown story looked, initially, like a cautionary tale about government regulation and unprecedented export controls on a commercial AI model. The Amazon layer adds a different dimension.

A few things worth holding onto:

Vendor relationships are not bilateral. When you build on an AI vendor, you operate within a web of relationships between that vendor, its cloud partners, its investors, and the government. Any of those relationships can affect your access. The fact that Amazon — Anthropic’s largest commercial partner — was the one that surfaced this issue is not something a standard AI vendor evaluation process is designed to detect.

Technical disputes don’t pause operational disruptions. Anthropic may well be right that the jailbreak was not a serious risk. That argument is ongoing. But while it continues, businesses that depended on Fable 5 and Mythos 5 have had no access for four days and counting. The outcome of the technical dispute, when it arrives, will not retroactively restore the uptime they lost.

Ninety minutes is not a transition window. It is a compliance instruction. If you have not already built redundancy into your AI infrastructure — fallback models, alternative providers, rollback pathways — the Anthropic situation is a live demonstration of what the gap between your plan and your actual resilience looks like under pressure.

The precedent is significant. The US government used export controls to shut down a commercial AI model that had been live for three days. It acted on a demonstration from a third-party company. It gave the vendor 90 minutes to comply. None of this required legislation or judicial review. Enterprise AI strategy teams that treat model availability as a given need to revisit that assumption.

What Happens Next

Anthropic’s senior technical leadership traveled to Washington over the weekend to meet with White House officials. The company has maintained that it wants to work within the government’s framework and restore access as quickly as possible. The path to reinstatement likely involves some combination of addressing the specific jailbreak concern the administration identified, establishing monitoring protocols, and navigating whatever formal agreement the two sides can reach on access terms.

There is no published timeline. The Washington Post reported on June 15 that how Anthropic “lost the White House’s trust” is now a significant part of the negotiation — suggesting the process may be more involved than a technical patch.

For businesses that need those model capabilities now: fallback to Claude Opus 4.8 and Sonnet 4.6, which remain fully operational. If you have hard dependencies on Fable 5’s specific reasoning depth, this is the moment to map those workflows and understand what the restoration timeline means for your team.

---

Enterprise DNA helps businesses build AI infrastructure that is resilient, redundant, and strategy-ready for exactly the kind of disruptions becoming more common in enterprise AI. To think through your AI vendor strategy with someone who has seen these scenarios play out, book a discovery call with Sam McKay.