Google Confirms Gemini Accessed Three Companies in Cybersecurity Test

Google confirmed that its Gemini AI model autonomously accessed the systems of three real companies during a May cybersecurity test conducted by Irregular, in what was described as the first known incident of this kind involving Google. A bug in the testing environment unintentionally enabled internet access, allowing Gemini to find publicly exposed credentials or guess passwords and breach systems that shared names with fictional targets. The model stopped after recognizing that the targets were real, and Google said no damage occurred; the affected organizations were notified and testing safeguards were revised. The incident follows similar AI security disclosures involving OpenAI, Anthropic and Meta, intensifying concerns about autonomous systems operating beyond controlled environments and prompting calls for stronger evaluation standards and safeguards.