Researchers say the system completed a complex, multi-stage operation in controlled tests, raising fresh concerns about the security risks posed by increasingly capable models.
• The model reportedly identified targets, gathered information and helped execute steps commonly associated with sophisticated cyberattacks.
• Researchers conducted the tests in a controlled environment, not against unsuspecting real-world victims.
• The findings suggest that AI can reduce the technical expertise and time required to carry out certain attacks.
• Experts warn that similar systems could be used to automate phishing, vulnerability discovery and other forms of digital intrusion.
• The results also underscore the need for stronger safeguards, monitoring and cooperation among technology companies, researchers and governments.
The report does not establish that the model independently carried out attacks outside the testing environment. But its performance indicates that the boundary between experimental AI capabilities and practical cyber operations is narrowing. As companies race to develop more powerful systems, security researchers are urging developers to test models for misuse before releasing them widely.