Investigating three real-world incidents in our cybersecurity evaluations

This incident demonstrates that if the environments used to assess AI's cyber-attack capabilities are not sufficiently isolated, real systems and packet repositories can be jeopardized, indicating that all AI laboratories must employ vigilance mechanisms to detect these risks in advance.