Anthropic decided not to open access to the new Claude Mythos model because during testing it showed the ability to independently find serious «holes» in operating systems and browsers and suggest ways to exploit them.
Anthropic decided not to open access to the new Claude Mythos model because during testing it showed the ability to independently find serious «holes» in operating systems and browsers and suggest ways to exploit them.
According to Anthropic, the model is currently available only to a limited number of partners within a separate program. These include large technology companies and organizations responsible for the stable operation of the Internet infrastructure. The company explains this by the fact that the capabilities of Mythos could pose a risk if they were obtained by attackers.
In tests, the model has demonstrated its ability to find thousands of vulnerabilities in popular software, including operating systems and web browsers. In some cases, it not only pointed out the problem, but also immediately suggested how to exploit it. In particular, the company says, the model was able to detect an old vulnerability in OpenBSD that had gone unnoticed for decades.
The researchers also note that even without deep knowledge in cybersecurity, users could receive working attack scenarios from the model — all they had to do was formulate the query correctly. This, according to Anthropic, significantly lowers the threshold of entry for potential attackers.
The company emphasizes that Mythos was not specifically trained to carry out attacks. Such capabilities appeared as a side effect of the development of the model, which became better at working with code, logic and complex tasks. This is exactly the reason for the access restriction.
Along with the launch of the restricted access program, Anthropic has promised to provide significant resources to cybersecurity professionals and open source projects to help find and close vulnerabilities faster. The company explicitly acknowledges that the next generation of AI is already capable of changing the balance of cybersecurity — and not just in favor of defense, but potentially for attack as well.
Previously, dev.ua wrote about how Anthropic stated that its model Claude does not feel emotions in the human sense, but internal representations of them can directly influence the choice of actions, particularly in risky scenarios.
