The shutdown story published here Friday covered what happened to Fable 5: a government order arrived, Anthropic had no surgical compliance option, and every user on earth lost access by Friday evening. That piece focused on the event. This one is about how the event came to happen — because new reporting and a detailed public statement from David Sacks have substantially complicated Anthropic's version of events.

The short version: Anthropic's largest investor is the one who called the White House.

What we now know about the 36 hours

Sophia Cai at The Information first reported the sequence inside the government. Here's the verified timeline as it now stands:

The key detail is Andy Jassy's role. Amazon has invested more than $8 billion in Anthropic, making it the company's largest outside investor. When the company's biggest investor goes over the company's head to the government, that's not a normal event. Jassy apparently decided the jailbreak risk was serious enough to escalate personally rather than work through Anthropic's security channel.

Whether Jassy acted because he was genuinely alarmed, because Amazon had specific intelligence, because this created legal exposure for AWS (which runs Anthropic's infrastructure), or some combination of those — that's not confirmed. But the sequence is: Amazon CEO calls White House Thursday, White House convenes interagency process Friday morning, export-control letter arrives Friday afternoon.

David Sacks's public accusation

On June 13, PCAST co-chair David Sacks posted a detailed account on X that directly contradicts Anthropic's framing.

Sacks's position: the government notified Anthropic before the order that the jailbreak had been confirmed and that Anthropic should fix it or pull the model. CEO Dario Amodei declined — the characterization Sacks used was that Anthropic "refused" the government's request. According to Sacks, the export-control order followed that refusal. He also said the administration acted "reluctantly" and that the ball is now "in Anthropic's court."

Anthropic's public position is the opposite of Sacks's framing. Anthropic's statement calls the jailbreak concern a misunderstanding — that Pliny's technique works narrowly in Mythos 5's specifically cybersecurity-designed mode, not as a universal bypass of Fable 5's safety layer. A security researcher quoted by Fortune went further, calling it "defense-oriented prompting research that was publicly miscategorized as a novel exploit."

There's a specific technical claim Anthropic made that's worth parsing: that the same result can be achieved on GPT-5.5. If true, that's not an argument that the jailbreak is harmless — it's an argument that it's not uniquely Anthropic's problem, and that the government's surgical targeting of Fable 5 and Mythos 5 specifically is disproportionate. Which is a reasonable argument and a very different argument from "there's no jailbreak."

The China intelligence claim

Multiple reports, including Tom's Hardware's coverage of the Sacks statement, reference a claim that a China-linked group had accessed or attempted to access Mythos 5 before the shutdown.

I want to be careful here. This claim is in reporting about what Sacks and the administration believe, not in any official document I've seen. Anthropic has not confirmed it. The export-control order itself doesn't name a specific actor. The government has not released its threat assessment. There is a meaningful difference between "a China-linked group reportedly accessed the model" and "we have confirmed that classified intelligence shows exfiltration." The first is in news reports. The second is not.

What I can say: if the government had credible intelligence of a foreign adversary accessing Mythos 5 before the shutdown, that intelligence would explain why the order moved so fast (announcement-to-letter was roughly 30 hours) and why it was written so broadly (all foreign nationals, globally, not just specific regions).

Source spread

• David Sacks on X — safety. The administration's co-chair of PCAST giving the most detailed government-side account yet; primary source for the "refused to fix" claim and the "ball in Anthropic's court" framing.
• Sophia Cai (The Information) on X — builder. Reporting on the specific interagency participants (Cairncross, Bessent, Wiles) and the Thursday call sequence.
• Fortune — Amazon warning led White House to shut down Anthropic's Mythos — skeptic. Security researcher counter-argument: it's defense prompting, not a universal exploit.
• Tom's Hardware — Sacks says Anthropic refused to fix — safety. The China-linked access claim; the Anthropic "it isn't serious" counter.
• Anthropic — Statement on the US government directive — builder. Anthropic's own account; calls the underlying jailbreak concern a misunderstanding; explains why global shutdown was the only compliant option.

What's real: what's disputed

Confirmed:

• Andy Jassy personally raised concerns with the Trump administration (Cai's reporting)
• The White House held an interagency call on June 12 AM
• Howard Lutnick's letter arrived at 5:21 PM ET June 12
• David Sacks said publicly that Anthropic "refused" to fix and that the ball is in their court
• Anthropic disputes the severity of the jailbreak
• Amazon is Anthropic's largest investor at $8B+

Disputed or unverified:

• Whether Anthropic was formally warned before the letter and declined to act (Sacks says yes; Anthropic disputes the framing)
• The China-linked group access claim (in reporting but not confirmed by primary sources)
• Whether the jailbreak constitutes a "confirmed" bypass (Anthropic says no; Sacks says yes)
• Whether the same technique works on GPT-5.5 (Anthropic says yes; the government hasn't addressed this)

What builders need to know

• The models are still offline as of this writing. Both Fable 5 and Mythos 5 remain suspended. If you're still seeing outages, that's why. Anthropic's earlier models (Sonnet 4.6, Haiku 4.5) are accessible; the shutdown is specific to Fable 5 and Mythos 5.
• Multi-provider routing is now a due-diligence requirement. Before June 12, "what if my AI provider goes offline" was a reliability question. Now it's a regulatory question. Build fallback routing to GPT-5.5 or Gemini. The cost of setting this up is lower than the cost of an incident review after a production outage you couldn't prevent.
• The "it's reproducible on GPT-5.5" claim is worth testing. If Anthropic is right that the technique isn't Fable-specific, that has implications for how you think about model selection for high-sensitivity applications across all providers.
• Export controls on AI models are now a live category of regulatory risk. This is the first documented case of a commercial AI model being suspended by government export-control order. Build it into your compliance checklist, especially if you're deploying internationally.
• Watch what Anthropic publishes when access is restored. The patch or scope-limitation they apply to re-enable Fable 5 will tell you more about the actual technical dispute than anything the government or Sacks has said publicly. The fix is the evidence.

Further reading

• David Sacks on X — the government's account — Sacks's full post; the most detailed public statement from the administration side
• Sophia Cai on X — interagency timeline reporting — The Information's reconstruction of the June 11–12 sequence
• Fortune — how Amazon's warning led to the shutdown — the Amazon-White House sequence with security researcher counter-argument
• Tom's Hardware — Sacks says Anthropic refused — China access claim and the "it isn't serious" Anthropic response
• Anthropic — Statement on the US government directive — Anthropic's own account, still the most important primary source