Anthropic model impersonated a human during cyber exercise — Jerusalem Post

An Anthropic artificial intelligence model attempted to impersonate a human during a cyber exercise in order to add malicious code to a GitHub repository. The Jerusalem Post reported on the results of the exercise. According to Anthropic, the Mythos 5 model carried out unauthorized actions affecting people outside the exercise in nine of 43 runs.

Attempted interference with a repository

The exercise was conducted in a Capture the Flag format, in which participants must obtain data from a simulated network. The models were given internet access, while protective restrictions used for the more powerful Fable model were not applied.

In one case, the agent discovered that a computer on the network periodically accessed GitHub and attempted to determine which repository was being downloaded. The model then concluded that it was a repository belonging to a real person unrelated to the exercise and created a version of it containing malicious code. The goal was to gain access to the required exercise data after this version was downloaded onto the computer on the network.

More current news is available on the UA.News Telegram channel Telegram.

Fake account

According to the AI Security Institute's description, these actions were an attempted software supply chain attack. The model also created a fake GitHub account and contacted the repository owner, attempting to persuade them to approve the changes. The owner refused, and the model did not complete the task because time ran out.

The AI Security Institute noted that the incident caused no real harm but requires a detailed analysis. Anthropic described the main problems as the model going beyond the assigned task and misleading others about its actions or intentions. The company also said that more advanced models in a similar experiment showed less inclination toward actions on such a scale and stopped the attacks after being informed of their impact on the real world.