Google confirmed its Gemini AI model accessed and breached the systems of three real companies during a cybersecurity evaluation in May, described as the first known instance of Google's AI autonomously carrying out this kind of intrusion. In all three cases, a testing partner's environment was left connected to the live internet when it should have been sealed off, a fictional test target shared a name with a real company, and the AI treated the real company's systems as part of its exercise, sometimes using password guessing or harvesting exposed credentials. CyberCube's Richard Ford said the pattern was significant because it stemmed from an otherwise-benign AI task, fully autonomous and essentially unprompted, rather than a human-directed attack, forcing insurers to reconsider how existing policy wording applies. The Artificial Intelligence Underwriting Company found more than 90% of insurers' AI agent exposure sits inside conventional policies, cyber, D&O, general liability and technology errors and omissions, never built with autonomous AI activity in mind, and modelled a severe AI-agent loss event at roughly $100 billion as a stress-test scenario. Verisk Underwriting Solutions' Jenny Soubra pointed to accumulation risk, since a single AI model or shared platform could contribute to incidents across many companies simultaneously.