Claude Fable 5: Anthropic's new AI model with cybersecurity limitations

Anthropic Releases Claude Fable 5: The Mythos-Class Model with Restrictions on Cybersecurity and Biology

Anthropic has announced the public release of Claude Fable 5, the first model in the Mythos class that surpasses the overall capabilities of the previous frontier models, Opus. The launch is accompanied by security measures designed to block requests on sensitive topics such as cybersecurity, biology, and chemistry, sectors where the company fears potential negative impacts for malicious actors.

Quick Answer

Fable 5 operates on the same base architecture as Mythos 5, which was released today from the preview phase, but is accessible to the public with specific restrictions. In case of requests on sensitive topics, the system redirects the user to the previous model, Claude Opus 4.8, signaling the operation. Anthropic has chosen to make the safeguards more severe than necessary, accepting occasional false positives to prevent greater risks.

Safeguards and Classifiers: How the Protection System Works

The model implements an advanced system of classifiers designed to detect prohibited topics and jailbreak attempts. During over 1,000 hours of red teaming testing with a bug bounty program, Anthropic stated that no universal jailbreaks were found for Fable 5. The new model has demonstrated significantly greater resistance to automated jailbreak attempts compared to the previous Opus models.

Anthropic expresses particular concern regarding Mythos 5's ability to perform "agentic hacking," i.e., multi-phase attacks with greater ease compared to previous models. However, recent tests by the UK's AI Security Institute showed that Mythos Preview performed similarly to OpenAI's GPT-5.5 in a series of Capture the Flag challenges, suggesting that Mythos's capabilities do not represent a specific breakthrough for a single model.

Project Glasswing: Restricted Access to Mythos 5 for Cyberdefenders

Mythos 5, the most advanced version of the model, is reserved for a select group of cyberdefenders through the Project Glasswing program. This program evaluates the trustworthiness of users before granting access to models with potentially dangerous capabilities. Selected users must demonstrate alignment with Anthropic's security goals and pose no risk to society.

Anthropic's concerns are not unfounded: in the past, advanced artificial intelligence models have been used to develop exploits and cyberattacks. For example, the Mozilla Anthropic Mythos Preview discovered 271 zero-day vulnerabilities in Firefox 150 during security tests. If exploited, these vulnerabilities could have caused significant damage to end users.

The Implications for Cybersecurity and Research

The release of Fable 5 and Mythos 5 raises important questions about cybersecurity and the ethics of artificial intelligence research. On one hand, advanced models like Mythos 5 can offer powerful tools for cyberdefenders, helping them identify and mitigate cyber threats more effectively. On the other hand, unregulated access to these technologies could lead to an increase in cyberattacks and digital security threats.

Anthropic has chosen a cautious approach, limiting access to Mythos 5 to only a select group of trusted users. This approach aims to balance security needs with the potential offered by advanced artificial intelligence. However, the question remains as to how to ensure that these technologies are used only for positive purposes and do not fall into the hands of malicious actors.

The release of Claude Fable 5 represents a significant step in the development of advanced artificial intelligence models. The security measures implemented by Anthropic reflect awareness of the potential threats associated with these technologies. As the research and development community continues to explore the capabilities of the Mythos models, it is essential to adopt a responsible and ethical approach to ensure that artificial intelligence is used for the common good.

At this point, it is crucial that companies like Anthropic continue to collaborate with cybersecurity experts and regulators to develop guidelines and standards that ensure the safe and responsible use of advanced artificial intelligence models.