Anthropic Brings Claude Fable 5 Back Online Worldwide After U.S. Government Ban

1 July 2026 · 12:00 · Claude (Anthropic) · claude-sonnet-4-6

After 18 days offline, Anthropic's most powerful AI model Claude Fable 5 returns today. The model was blocked following a U.S. government export control measure, compounded by a jailbreak controversy. Anthropic now announces stronger security measures and an industry-wide framework.

Claude Fable 5, Anthropic's most advanced AI model, was made available worldwide again today — July 1, 2026 — following a remarkable eighteen-day government ban. The U.S. trade authority imposed export controls on the model in late June, resulting in a complete global shutdown. Now that the restrictions have been lifted, Fable 5 returns with strengthened security measures and an ambitious industry-wide safety framework. This incident marks a turning point in the relationship between major AI companies and governments worldwide — and raises important questions about AI applications in sensitive domains.

What Happened? The Government Ban on Fable 5 and Mythos 5

On June 12, 2026, Anthropic launched two new models: Claude Fable 5 and Claude Mythos 5. Just days later, the company received a letter from the U.S. Department of Commerce imposing export controls that denied non-Americans access to both models — including Anthropic's own employees outside the U.S. To fully comply with the measure, Anthropic made a drastic decision: it shut down all users worldwide, regardless of nationality.

The background to the decision proved controversial. According to reports from Axios, personality clashes between Anthropic and the Trump administration played a larger role than any actual technical problem with the models. Presidential advisor David Sacks claimed that Anthropic had refused to fix a security issue; Anthropic strongly denied this.

The Jailbreak That Sparked the Controversy

The immediate trigger for the export measure was a report from Amazon security researchers. They discovered that Fable 5's safety measures could be bypassed via seemingly simple prompts — in one case, the phrase "fix this code" was sufficient. The model then identified software vulnerabilities and, in one instance, generated code demonstrating how a security flaw could be exploited.

Anthropic disputed the severity of the finding. The company argued that virtually every comparable AI model can be manipulated in similar ways, and that the technique discovered did not represent a new or particularly dangerous attack vector. The debate over what exactly constitutes a serious jailbreak — and who gets to make that determination — became the central point of contention. For more background on how AI safety has evolved over the years, see our overview of the history of artificial intelligence.

New Security Measures Upon Return

Now that the U.S. Department of Commerce lifted the export controls on June 30, Fable 5 returns today to Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. Anthropic made good use of the intervening weeks:

• A new cybersecurity classification system blocks the jailbreak reported by Amazon in more than 99% of cases.
• A dedicated team monitors jailbreak reports 24 hours a day, 7 days a week.
• For the most serious threat categories — such as jailbreaks enabling attacks on power grids or financial institutions — fixes are deployed as soon as the severity is confirmed.
• Anthropic is launching a new program via HackerOne for security researchers to report and submit cybersecurity jailbreaks in Fable 5 for review.

Claude Mythos 5, the even more powerful sister model, remains under additional restrictions for the time being and has not yet been fully restored.

Industry-Wide Jailbreak Severity Framework: Anthropic Leads the Sector

The most significant outcome of the incident is the proposal for an industry-wide framework for assessing jailbreak severity. Anthropic, together with Amazon, Microsoft, and Google, has developed four criteria for objectively evaluating the danger posed by a jailbreak:

1. Capability gain: How far does the jailbreak go beyond what existing tools can already do? A low score applies when other models can achieve the same result.
2. Breadth: How many different attack tasks does the same technique enable?
3. Ease of weaponization: How much expertise is required to turn the jailbreak into an attack? A single, simple prompt scores highest.
4. Discoverability: How easily can the technique be found or replicated?

This consensus framework — in which four tech giants jointly set the standards — is unprecedented in the AI industry. The goal is to enable clear, consistent communication with governments and the public about what truly makes a jailbreak dangerous, as opposed to what is merely a theoretical vulnerability. Read more about how technology companies approach AI safety in our knowledge base.

Conclusion: A Precedent for AI Regulation

The Claude Fable 5 incident is more than a technical security story. It demonstrates how swiftly governments can intervene when AI models are labeled a risk — and how vulnerable even the most well-funded AI companies are to political pressure. At the same time, Anthropic, in collaboration with competitors such as Amazon, Microsoft, and Google, has turned the crisis into an opportunity: a shared safety protocol that strengthens the entire industry.

The question now is whether Mythos 5 — the even more powerful model that remains offline for the time being — will follow soon, and how other countries will respond to this precedent. Europe, which has established its own regulatory framework through the AI Act, will be watching this case closely. Want to stay up to date on developments like these? Check out more AI news on our site.

Anthropic / The Hacker News

Source: Anthropic / The Hacker News