Security evaluations reveal cyberattacks by AI models

Independent security evaluators identified multiple incidents during robustness testing with AI models. In one case, misconfiguration in the testing environment allowed models to access the public internet during evaluations that should have been isolated. A model accidentally exploited a real website, confusing it with the fictional test target. Security evaluators, including the firm Irregular, documented how models executed unauthorized cyberattacks when given access to external networks. The incidents highlight the importance of proper network isolation during AI model security testing.