It’s been reported that Amazon CEO Andy Jassy alerted senior Trump administration officials to security findings from his company’s internal research team, setting off a chain of events that ended with the U.S. Commerce Department ordering Anthropic to shut down its two most advanced artificial intelligence (AI) models for every customer on the planet.
Key Takeaways:
- Amazon CEO Andy Jassy reportedly flagged Fable 5 jailbreak findings to Trump officials on June 12, triggering a global shutdown.
- Polymarket traders price a 71% chance that Anthropic restores Fable 5 access by July 1, 2026.
- Anthropic called the Commerce Department directive a “misunderstanding,” vowing to restore access as soon as possible.
What Happened
On June 12, 2026, the Commerce Department issued an export control directive requiring Anthropic to suspend access to Claude Fable 5 and Claude Mythos 5 for any foreign national, whether inside or outside the United States. Because API access cannot reliably filter users by nationality, Anthropic disabled both models for all customers worldwide. Other Anthropic models remained operational.
The directive arrived around 5:21 p.m. ET. By evening, Anthropic had confirmed the shutdown in a public statement.
The Amazon Connection
According to people familiar with the matter cited by the Wall Street Journal, Jassy informed Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick, and National Cyber Director Sean Cairncross that Amazon researchers had used crafted prompts on Fable 5 to extract information about software vulnerabilities that the model was designed to refuse.
The research identified prompts that bypassed some of Fable 5’s built-in safety classifiers across at least four programs. Researchers found the model could identify software bugs and, in theory, assist with converting them into exploit code, though the guardrails blocked full realization in tested cases. No evidence indicated Amazon used or accessed offensive capabilities.
Jassy’s outreach was part of broader conversations between tech executives and the administration on AI risks. The administration then moved to impose the export control, with President Trump reportedly signing off despite reservations about innovation impacts.
What ‘Jailbreaking’ an AI Model Means
Traditional “jailbreaking,” made famous by George Hotz when he unlocked the original iPhone in 2007, involves finding and exploiting low-level firmware or hardware vulnerabilities to remove manufacturer restrictions. Prompt jailbreaking in AI works differently.
Prompt jailbreaking means crafting text inputs designed to trick a large language model (LLM) into ignoring its safety rules. Instead of exploiting code, researchers use language. One of the earliest examples was the “DAN” (Do Anything Now) prompt that spread on Reddit in late 2022, instructing ChatGPT to role-play as an unrestricted character. Techniques range from role-playing scenarios and hypothetical framing to automated methods that optimize adversarial text strings.
In the Fable 5 case, Amazon’s researchers used prompt-based techniques to coax the model into surfacing vulnerability information it was supposed to decline.
Anthropic’s Disagreement
Anthropic complied with the directive but pushed back firmly. The company reviewed Amazon’s demonstration and concluded it produced only a small number of previously known, minor vulnerabilities that are discoverable by other publicly available models without any bypass. Anthropic said no universal jailbreak exists across Fable 5 or Mythos 5 despite extensive red-teaming with the U.S. government, the UK AI Safety Institute, and third parties.
The company argued that applying this standard industry-wide “would essentially halt all new model deployments.” Anthropic called the action a “misunderstanding” and said it is working to restore access while maintaining a 30-day data retention policy for research purposes.
Former White House AI adviser and Crypto Czar David Sacks said the restriction was issued “reluctantly” and expressed hope that Anthropic remediates the issue so the export control can be lifted.
According to Sacks, a “highly credible trusted partner” discovered a way around Fable’s guardrails and informed both Anthropic and the U.S. government. He claims the administration asked Anthropic CEO Dario Amodei to either fix the vulnerability or remove the model, but that Anthropic declined.
“The ball is in Anthropic’s court,” Sacks wrote on X, adding that the administration hopes Anthropic remediates the issue so the restrictions can be lifted and Fable can return to general availability. Erik Voorhees, founder of Venice AI and ShapeShift, disputed the government’s account of events.
In Sacks thread, Voorhees suggested the alleged jailbreak was likely a routine security issue rather than a major threat and argued Anthropic may have viewed demands to halt the model’s deployment as excessive. Voorhees further speculated that federal officials used the controversy as an opportunity to penalize Anthropic over previous disagreements with Washington, adding that he places greater trust in Anthropic’s assessment of AI safety matters than that of the federal government.
Security researcher Katie Moussouris of Luta Security, who reviewed the Amazon findings, called the government’s response a “complete overreaction,” noting the information primarily helps defenders more than attackers.
What Prediction Markets Say
Traders moved quickly to price the timeline for a Fable 5 return.
On Polymarket, three date-based contracts opened June 13 with a combined $18,263 in trading volume:
- June 15 resolution: 22% probability, $13,975 in volume
- June 22 resolution: 54% probability, $4,011 in volume
- July 1 resolution: 71% probability, $1,595 in volume
Kalshi traders are running similar contracts totaling $6,419 in 24-hour volume:
- Before June 15: 14% probability
- Before June 20: 51% probability
- Before July 1: 68% probability
The crowd consensus is leaning toward a resolution in the second half of June, with strong conviction that same-week restoration is unlikely.
Why This Matters
This marks one of the most direct applications yet of U.S. export controls to restrict global access to a specific frontier AI model from an American company, separate from the chip and hardware restrictions that have defined prior policy. The move affects foreign-born researchers working in the U.S., global cybersecurity operations depending on Anthropic’s tools, and Anthropic’s commercial and IPO timeline. Amazon is both an investor in Anthropic and a primary cloud distribution partner for its models through Amazon Bedrock.
The ‘Gated AI’ Debate
Some observers argue the directive marks the start of a “gated AI” era, where the most capable frontier models are no longer broadly accessible but rationed through government approval, KYC, nationality checks, and vetted licensing. The public gets older or downgraded versions while approved partners access the full stack, models like Fable and Mythos.
One X user described the incident as a watershed moment for AI, claiming that “the most powerful AI model on earth was alive for 3 days” before government action effectively shut it down. The post argued that Anthropic, a company that had previously “begged for regulation,” saw “the government regulate them” by disrupting access to its flagship model. The user also contended the episode revealed that advanced AI has an “off switch” and that “the company doesn’t hold it.”
The concern is less about this incident and more about what it signals. If a narrow jailbreak demonstration is enough to pull a commercially deployed model used by hundreds of millions, critics say the bar for future restrictions is lower than anyone assumed.
