Anthropic unveils Claude Mythos 5 and Fable 5, with built-in protection against cyberattacks

Anthropic has introduced two new versions of Claude: Mythos 5 for trusted organizations and Fable 5 for the general public. The second version was designed specifically with mechanisms that make it harder to use in cyberattacks.

Split by purpose

Anthropic is separating its AI models according to target audience and potential risk. Claude Mythos 5 is intended for trusted partners, including cybersecurity companies, research organizations and enterprise customers who need the full capabilities of the language model without restrictions.

Claude Fable 5 is available openly to the general public. The company built special restrictions into this version to prevent the model from being used in cyberattacks and other malicious purposes. The strategy lets Anthropic offer a powerful AI tool while minimizing its direct responsibility for potential misuse.

Protection built into the architecture

The key difference with Fable 5 is that its protection mechanism does not rely merely on instructions or API-level prohibitions (which are easy to bypass). It rests on changes to the architecture and training process of the model itself. Anthropic believes this approach is more reliable and durable than application-level control systems.

• Restrictions built in during model training
• Architectural changes that hinder certain classes of tasks
• Public availability without licensing barriers, but with restrictions

Unlike a simple filter that catches malicious requests in real time, built-in restrictions change the model's behavior at a fundamental level. This means that even when someone tries to circumvent them, the system will be less useful for cyberattacks.

Growing pressure on the industry

Anthropic's decision reflects mounting pressure from regulators, politicians and the public, who demand that AI companies take responsibility for potential harm. Amid debate over whether large language models can be used in cyberattacks, releasing a "safe" version looks like a strategic move.

Other companies, including OpenAI and Google, also face pressure and are developing their own approaches to control. But the industry still has no single standard. Anthropic's two-tier approach looks pragmatic: give trusted partners full access, and offer the public a limited but safe version.

What this means

Anthropic is showing that AI safety is starting to be treated as a separate product line rather than an add-on module. This may become an industry trend: instead of one model for everyone, companies will offer several versions with different levels of restriction.

Questions remain, however, about how effective the built-in restrictions are in practice, and whether users will accept limitations in exchange for a promise of safety.