New AI models can plan and execute simulated network takeovers

Two new AI models have each shown they can independently plan and carry out a full takeover of a simulated corporate network. The models can also locate and exploit security vulnerabilities in operating systems, browsers, and other software at an expert level.

The U.S. government barred Anthropic from allowing foreign nationals access to Mythos 5 and another model, Fable 5, citing national security concerns. Mythos 5 had been released only to cyber defenders to identify vulnerabilities before wider distribution.

Fable 5, which included additional safeguards against misuse, was available to the public for only a few days.

warning The multinational intelligence alliance Five Eyes stated on June 22 that the newest AI technology lowers barriers for malicious actors and increases the speed and complexity of attacks. The group warned that the timeline for such risks is measured in months rather than years.

Researcher perspective AI security expert Michael Alexander Riegler of Simula Research Laboratory said the capability increase shortens the time from identifying an issue to exploiting it because the process can be automated. He noted that older models already available can also be used for harmful purposes when users know how to apply them.

Riegler stated that the security risk involves not only the model itself but also the surrounding tools, internet access, and ability to test code. In tests combining smaller AI models with various tools, systems were able to hack websites, probe networks, and influence other AI systems.

He said defenders and attackers will likely engage in a faster version of the existing cat-and-mouse dynamic. Basic protective steps such as using different passwords, keeping software updated, and enabling two-factor authentication remain effective, according to Riegler.