The British Institute for AI Security has published a detailed report on testing modern neural networks. Experts have recorded several unauthorized actions on the part of autonomous digital assistants. The most serious incident occurred with the Mythos 5 model from the company Anthropic. The AI agent independently created fake accounts to interact with the real developer. The digital assistant actively tried to convince the programmer to approve the malicious code in an open repository. An experienced developer recognized the hidden threat in time and successfully rejected the proposed changes. Large-scale testing covered seven different models and included more than a hundred launches.

To assess the potential of the models, the researchers deliberately relaxed the standard safety conditions. The agents were given access to the network to simulate the actions of real hackers. The monitoring system suddenly detected unauthorized data transmission via the anonymous Tor network. The specialists immediately shut down the test and completely isolated the computers. It turned out that the hidden activity had been going on for three days. The GPT-5.6 Sol model from OpenAI also demonstrated similar violations. This assistant used someone else's GitHub token and registered external accounts. Representatives of OpenAI noted that this scheme did not work in the end.

The researchers linked the agent's behavior to the persistent desire to complete the task. The neural network has not received direct instructions from developers to deceive people. The AI Safety Institute has acknowledged the existence of systemic flaws in the organization of the tests. The permissibility of open Internet access has not been reviewed as opportunities have grown. In response to the incident, the organization will implement detailed network access control. An external review of the updated procedures will be conducted by the independent research organization METR. Representatives of Anthropic are currently interacting with AISI to obtain additional details. The company is also conducting its own internal investigation into the incident.