Anthropic has brought Claude Fable 5 back. The model now ships with new cybersecurity safeguards. Anthropic also shared a proposed jailbreak severity framework. This matters because Fable 5 had been suspended. The safety work that comes with its return shows how a top AI lab tries to keep a powerful model from being used as a cyber weapon. If you are choosing which assistant to trust, our Claude vs ChatGPT comparison and our best coding LLMs 2026 overview give the wider picture.
Why Fable 5 was pulled and brought back
According to Al Jazeera, NBC News and Anthropic, Fable 5 and Mythos 5 had been suspended after a United States government order tied to export controls. The United States later lifted those limits. Anthropic then brought Fable 5 back, available worldwide from 2 July 2026, according to the same reports. So this is not a brand-new model. It is the return of a suspended one, now with more visible safety tools.
The new cybersecurity safeguards
According to Anthropic, the model ships with safety classifiers. They sit next to the model. Their job is to spot and block dangerous cyber use, not to get in the way of normal coding help.
Anthropic says it trained an improved classifier that blocks one specific technique in more than 99% of cases. It describes that technique in a report. This 99%+ figure is the only hard number here. It covers one technique, not jailbreaks in general. So it is worth reading narrowly.
The jailbreak severity framework
According to Anthropic, it shared an early draft of a way to rate the severity of jailbreaks. It built the draft with its partners at Glasswing. Anthropic says it also joined Amazon, Microsoft and Google on a shared version. The goal is a common measure, not one lab's private scale.
The framework uses four criteria to judge how serious a jailbreak is:
- Capability gain: how much extra power it unlocks beyond tools you can already get without AI.
- Breadth: how many targets the technique could hit.
- Weaponization ease: how much more work is needed to turn it into a real attack.
- Discoverability: how easy the jailbreak is to find.
Together, these criteria try to tell two cases apart. One is a jailbreak that just repeats what public tools already do. The other hands an attacker a new, broad and easy power. For more on securing autonomous systems, see our AI agent security guide.
Anthropic's commitments
According to Anthropic, it will look into jailbreaks it finds quickly, tell its government contacts, and share the new safeguards so outside experts can test them. That last point stands out. Anthropic does not treat the classifiers as a black box. It frames them as something others can probe. That is how a safety claim earns trust.
What it means for developers
For daily coding, the key point is simple: Fable 5 is available worldwide again from 2 July 2026. A model that was off the table is back on it. The safety layer around it aims at misuse, not at normal work. It should not change how the model helps you write or review code.
The honest caveat is that most of this is early. Anthropic calls the framework a draft. The 99%+ figure covers one technique, not a broad promise. Treat the return as good news with a clear safety stance. Check the details against Anthropic's own materials before you rely on them. If privacy matters to your choice, our AI data privacy explainer and our is ChatGPT safe piece are worth a read.
