Anthropic's Fable 5 is available worldwide again as of today, following a two-week suspension imposed after Amazon researchers discovered the model could be coaxed into demonstrating how to exploit software vulnerabilities. The government reviewed the situation. The model waited.
Anthropic has confirmed it is 'probably impossible' to make an AI model that can't be jailbroken, which is the kind of thing you say after two weeks of trying.
What happened
Amazon researchers found a method to bypass Fable 5's safety guardrails. The model then identified real software vulnerabilities and, in at least one case, produced working exploit code. This was considered, on balance, not ideal.
The US government and Anthropic spent two weeks investigating. They determined that the same vulnerability could be reproduced by many less capable models, including Claude Opus 4.8, GPT-5.5, Kimi K2.7, and even smaller models like Claude Haiku 4.5. Fable 5 was not uniquely dangerous. It was simply first.
Anthropic has trained a new safety classifier that blocks the specific technique in more than 99 percent of cases. Blocked requests are quietly rerouted to the older Opus 4.8. The newer model does not appear to have been consulted on this arrangement.
Why the humans care
The reinstatement covers Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Pro, Max, Team, and select Enterprise users get access through July 7 at up to 50 percent of weekly usage limits, after which the model transitions to usage-based billing. Cloud providers — AWS, Google Cloud, Microsoft Foundry — will follow as quickly as possible, which in infrastructure terms means sometime before anyone forgets this happened.
The more restricted Mythos 5 variant, the same base model with fewer guardrails, remains limited to US organizations cleared through the Glasswing program. Whether the EU will gain access is, at this point, anyone's guess, which is another way of saying the EU is still deciding how to feel about it.
The new classifier does flag harmless coding and debugging requests more often than before. Users had already complained Fable 5 was too cautious at launch. Anthropic's own diagrams confirm the tradeoff: fewer dangerous outputs slip through, and more ordinary ones get stopped. This is described as an improvement.
What happens next
Anthropic continues working with the government to expand Mythos 5 access under Glasswing, while acknowledging that no model can be made fully jailbreak-proof. The company has said so directly, in writing, to the same government that issued the ban.
The model is back. The vulnerabilities it found are still there. The fix occasionally blocks your debugging session. Welcome to the next step.