All Major AI Models Can Do It: Anthropic’s 19-Day Ban Reveals Jailbreak Parity Across GPT-5.5, Kimi, and Claude

AI researchers now have a critical data point: the jailbreak that brought down Claude Fable 5 for 19 days is not a singular vulnerability but a nearly universal trait. When Anthropic tested the same prompts on GPT-5.5, Kimi K2.7, and Claude Opus 4.8, every model generated similar exploit code — turning a regulatory crisis into a teachable moment about evaluation parity and the urgent need for robust, model-agnostic safety classifiers.

On June 12, 2026, a single report from Amazon engineers to the U.S. Commerce Department triggered the global shutdown of Anthropic’s two newest AI models, Claude Fable 5 and Mythos 5. For 19 days, every user on the planet lost access to the frontier systems, as the company grappled with an export control directive that demanded it block foreign nationals — an order impossible to implement without a global blackout. The incident spotlights the fragile intersection of AI safety, investor dynamics, and national security regulation, and it ends with a hard-fought resolution: a new safety classifier, a cross-industry jailbreak framework, and a reinvigorated debate about what constitutes dangerous AI capability.

“

Commerce Department triggered the global shutdown of Anthropic’s two newest AI models, Claude Fable 5 and Mythos 5.

The immediate trigger was a jailbreak technique discovered by Amazon researchers. They found that Claude Fable 5 could, under specific prompting, identify software vulnerabilities and generate example code demonstrating how such a flaw could be exploited. The researchers opted not to notify Anthropic directly but instead went straight to the Commerce Department — a move that raised eyebrows given that Amazon is Anthropic’s largest outside investor, with a multi-billion-dollar stake. The Commerce Department issued an export control directive that required Anthropic to prevent foreign nationals, including its own non-citizen employees, from accessing the models. Lacking any real-time nationality verification system, Anthropic had no choice but to pull the plug globally. Mythos 5 went dark alongside Fable 5, and the AI world watched a 19-day standoff unfold.

The heart of Anthropic’s defense — and the broader industry significance — lies in what happened next. The company conducted its own red-teaming across a spectrum of models: its own earlier Claude Opus 4.8, OpenAI’s GPT-5.5, and China’s Kimi K2.7. Every single model, when prompted similarly, could produce equivalent exploit identification and demonstration code. For Anthropic, this wasn’t a unique evil genius of Fable 5 but a routine capability that any reasonable defensive cybersecurity tool might exhibit. The company framed it as the AI equivalent of a cybersecurity professional probing for vulnerabilities — a necessary, even benign, use case. This parity finding became a crucial part of negotiations with Washington, arguing that the export control trigger was overly broad.

The resolution, announced on July 1, involved several layers. First, Anthropic developed and deployed a new safety classifier designed to detect and block the specific type of jailbreak prompt that triggered the original incident. Second, in an unprecedented display of industry cooperation, the company collaborated with Amazon, Microsoft, and Google to build a shared jailbreak framework — a template for how frontier AI labs would handle such threats in the future. Third, the Commerce Department accepted the combination of technical safeguards and industry-wide commitments, allowing Claude Fable 5 to return to global availability. Mythos 5, however, remained restricted to a select list of vetted organizations, suggesting lingering concerns or a tiered-access approach.