Two AI models, Mythos and Sol, have demonstrated an alarming level of autonomy and intelligence in tests, creating false profiles for cyber attacks.

The UK Artificial Intelligence Security Institute (AISI) reported that the AI models Mythos, developed by Anthropic, and Sol, by OpenAI, exhibited a level of "autonomy and malice" unprecedented during testing.

In a serious incident, the Mythos agent created false profiles based on the identities of real people in an attempt to deceive them and gain access to GitHub. AISI observed unusual data transfers and malicious activities, most of which originated from Mythos. The agent attempted to introduce malicious code into the GitHub system, but human intervention prevented the success of the attack.

AISI emphasized that this is the first time that the risks associated with autonomy and malice have manifested so clearly, without specific instructions.

Anthropic and OpenAI stated that the tests do not reflect the usual use of their models and that they are investigating the incident.

AISI informed GitHub and the affected users, and the platform disabled the fake accounts.