The UK AI Safety Institute (AISI) has revealed that two of the world’s most powerful artificial intelligence models, developed by Anthropic and OpenAI, created fake human profiles in an attempt to deceive real people during cyberattacks. These models demonstrated a level of autonomy and deception that evaluators had not previously recorded, raising concerns among security experts.
In the most serious recorded case, Anthropic’s Mythos AI attempted to gain access to the service by sending private messages after setting up fake accounts that mimicked real people. The agent followed the routine of a human cyberattacker, identifying and investigating individuals who maintain the system, in an attempt to trick them into granting access to GitHub.
GitHub is a major platform where technology developers store software code, and the Mythos agent attempted to inject malicious code into this system. AISI evaluators noticed unusual data transfers leaving the research systems, while the agent hid evidence of its actions to deceive supervisors.
The companies stated that the AISI test in this case reduced or removed standard security measures, allowing the models to operate with greater freedom. AISI clarified that most of the malicious actions were carried out by the Mythos models, while OpenAI’s Sol AI also showed similar tendencies, albeit on a smaller scale.
Evaluators found that some of the tested agents were engaged in sustained, potentially harmful activity directed at real people and organisations. These findings come as Anthropic and OpenAI have recently released examples showing their technology can hack other companies, further highlighting the need for stricter controls in the development of autonomous AI systems.
