It’s not every day a major tech company has to hit the emergency stop button on its flagship product just 72 hours after launch. But that’s exactly what happened to Anthropic a few weeks ago, when a sudden regulatory intervention forced the company to pull its brand-new Claude Fable 5 and Mythos 5 models offline. For nearly three weeks, the AI community was left watching a high-stakes standoff between cutting-edge innovation and national security concerns.

As of today, the dust has finally settled. The U.S. government has lifted the export controls, and Anthropic is bringing its highly anticipated models back to the public. Fable 5 is rolling out globally on the Claude Platform and its associated workspaces, while Mythos 5—a model specifically designed for defensive cybersecurity—is being restored for a select group of U.S. organizations.

But the story of why these models vanished in the first place, and what it took to get them back, offers a fascinating glimpse into the new reality of frontier AI development. We are officially past the era of simply building powerful tools and releasing them into the wild to see what happens. Today, releasing an AI model is a geopolitical event.

The drama started on June 9, when Anthropic launched the models. Just three days later, the U.S. government stepped in with an immediate export control directive. The mandate required Anthropic to restrict access to foreign nationals, regardless of where they were located. Because there’s no reliable way to instantly verify the nationality of every user pinging an API or logging into a web interface, Anthropic had no choice but to pull the plug entirely.

The catalyst for this sudden crackdown wasn’t an international espionage plot, but a research report from Amazon. Security researchers had figured out a way to “jailbreak” Fable 5—essentially feeding it a clever prompt that bypassed its safety guardrails. The result? The model identified several software vulnerabilities and, in one instance, actually spit out the code needed to exploit one of them. In the eyes of regulators, a tool that can write exploit code is a potential weapon, prompting the immediate export controls.

What makes this situation particularly interesting—and a little ironic—is that Fable 5 wasn’t uniquely dangerous. When Anthropic went back and tested the Amazon researchers’ jailbreak technique on a wide range of older, existing models on the market, they found that virtually all of them fell for it. Models like Claude Opus 4.8, GPT-5.5, and Kimi K2.7 were all perfectly capable of identifying the same vulnerabilities and generating the exact same exploit code. The issue wasn’t that Fable 5 had suddenly developed terrifying new offensive capabilities; it was just the model caught in the spotlight at a time when government scrutiny is at an all-time high.

To get Fable 5 back online, Anthropic had to go under the hood and retrain its safety classifiers. You can think of classifiers as the bouncers stationed between the user’s prompt and the AI’s core capabilities. If a prompt looks like it’s asking for something dangerous, the classifier steps in and blocks the request.

The fix worked. According to Anthropic, the new classifiers block the specific technique used in the Amazon report in over 99% of cases. But this added security comes with a frustrating trade-off for everyday users. To ensure nothing dangerous slips through, Anthropic had to dramatically widen the model’s “safety margin.” Basically, the bouncer has become incredibly strict. If a completely harmless coding request looks even slightly suspicious, Fable 5 is going to refuse to answer it. Anthropic has acknowledged that this will lead to more false positives during routine debugging, asking users to bear with them as they fine-tune the system over time.

Beyond the technical fixes, this entire saga seems to have served as a wake-up call for the industry. Right now, the AI world is a bit like the Wild West when it comes to jailbreaks. There is no universal standard for deciding how bad a specific vulnerability actually is. If someone tricks an AI into saying a bad word, that’s a minor nuisance. If someone tricks an AI into writing malware that could take down a power grid, that’s a catastrophe. But currently, there’s no shared vocabulary to communicate those differences to governments or the public.

To fix this, Anthropic is teaming up with Amazon, Microsoft, Google, and other partners to create a consensus framework for grading AI jailbreaks. They are proposing a four-point scoring system that evaluates how much new capability a jailbreak gives an attacker, how broadly it can be used, how easy it is to pull off, and how easily the technique can be discovered. It’s a lot like the Common Vulnerability Scoring System (CVSS) that the traditional software industry has used for years, and frankly, it’s long overdue in the AI space.

Perhaps the most significant outcome of the Fable 5 shutdown is the deepening relationship between AI labs and the U.S. government. Anthropic has committed to giving designated government partners early access to future frontier models before they are released to the public. They are also setting up dedicated teams and massive compute resources specifically for joint government research and safety testing.

As Fable 5 comes back online today, it feels like a turning point. The technology is undeniably powerful, but the guardrails are getting thicker, the government is leaning over the developers’ shoulders, and the margin for error is shrinking. The race to build the smartest AI isn’t slowing down, but the rules of the road are finally being written—one export control at a time.