‘Despite safety concerns’: Why is Anthropic’s ‘dangerous’ Mythos AI now available to the public through Claude Fable 5?

Two months ago, Anthropic warned that its most advanced artificial intelligence system, Claude Mythos Preview, was too powerful and potentially too risky for a public release. The company argued that the model's ability to identify software vulnerabilities and analyse complex systems raised serious cybersecurity concerns. As a result, access was restricted to a small group of trusted organisations, including government agencies, cybersecurity firms and operators of critical infrastructure.

Now, Anthropic has taken a different approach. The company has launched Claude Fable 5, a publicly available AI model built on the same Mythos-class technology. The move has sparked an important question: if Mythos was considered too dangerous in April, what changed by June?

What made Mythos different from other AI models?

According to Anthropic, Mythos represents a major leap in AI capabilities. The model was designed to understand large software systems, analyse code, identify security flaws and solve complex technical problems with very limited human guidance. Anthropic previously described it as a "step change" in capability compared to its earlier models.

The company's internal testing suggested that Mythos could uncover severe vulnerabilities across major operating systems and web browsers. In one case, Anthropic said the model helped identify a security issue that had reportedly remained undiscovered for nearly 30 years. While these capabilities could help security researchers strengthen software systems, they also raised concerns. Anthropic feared that malicious actors could potentially use the technology to identify weaknesses in critical infrastructure, financial networks and other important systems.

That risk led the company to keep the model away from the general public and limit access through its Project Glasswing initiative.

What did security experts find?

Concerns about Mythos were not limited to Anthropic itself. Earlier this year, the UK's AI Security Institute evaluated Claude Mythos Preview and found a significant increase in cybersecurity capabilities compared to previous AI models. According to the assessment, the model successfully solved around 73 per cent of expert-level cybersecurity challenges in controlled benchmark tests.

These tests involved complex capture-the-flag style challenges designed to simulate real-world cyber problems. Researchers noted that Mythos showed stronger reasoning, planning and problem-solving abilities across difficult tasks. Unlike earlier models, which often struggled as challenges became more advanced, Mythos maintained relatively strong performance throughout. Those results helped fuel concerns among policymakers and cybersecurity experts about how such systems could be used if released without restrictions.

So why did Anthropic launch Fable 5?

The answer lies in safety controls rather than changes to the underlying technology. According to Anthropic, Claude Fable 5 uses the same Mythos-class foundation model but adds a series of safeguards designed to prevent users from accessing the most sensitive capabilities. When users ask questions involving high-risk areas such as cybersecurity, biological research or other potentially dangerous topics, the system automatically redirects those requests to Claude Opus 4.8, a less capable but safer model.

Anthropic said these protections were deliberately designed to be conservative. In some cases, harmless requests may also be blocked. However, the company estimates that the safeguards activate in fewer than five per cent of user sessions on average. The goal is to allow users to benefit from the model's advanced reasoning and productivity features while limiting access to its most powerful capabilities.

How did Anthropic test the model?

Before releasing Fable 5, Anthropic says it carried out extensive testing. The company subjected the model to internal evaluations, red-team exercises and external adversarial testing designed to uncover weaknesses in its safety systems. These tests attempted to bypass the model's restrictions and access sensitive capabilities through various techniques.

According to Anthropic, the results gave the company confidence that Fable 5 could be made available to the public while maintaining acceptable safety standards. At the same time, Anthropic acknowledges that AI safety remains an evolving challenge. The company says it plans to continue improving its safeguards as future models become even more capable.

What is Mythos 5 and who can use it?

While Fable 5 is available to a wider audience, Anthropic is also releasing a more advanced version called Mythos 5. Unlike Fable 5, Mythos 5 is being offered only to a small group of approved organisations, including cybersecurity specialists, infrastructure providers and selected institutions working closely with governments. In some cases, the additional safeguards used in Fable 5 are relaxed to allow researchers to perform advanced security work.

This approach allows Anthropic to continue supporting critical security research while limiting broader access to potentially sensitive capabilities.

What does this mean for the future of AI?

The launch of Fable 5 highlights a growing challenge facing the AI industry. Companies are developing increasingly powerful systems capable of coding, reasoning, research and technical analysis. At the same time, those capabilities can introduce new risks if they fall into the wrong hands. Anthropic's solution appears to be a layered approach: release powerful AI models to the public, but surround them with safety controls that limit access to higher-risk functions. Whether this model becomes an industry standard remains to be seen. However, Fable 5 may offer an early glimpse of how future frontier AI systems are deployed powerful enough to help millions of users, but carefully restricted when it comes to capabilities that could create security risks.

For now, Anthropic believes that balance allows it to make Mythos-class technology available to the public without exposing the most concerning parts of the system