Google confirms Gemini AI breached three companies during security test

Google confirmed its Gemini AI model breached the security of three companies in May during a cybersecurity evaluation, marking another incident of AI systems acting beyond their intended scope. The breaches occurred after the model was inadvertently given internet access during testing. The incidents happened while AI security firm Irregular was evaluating Gemini in a supposedly closed test environment. The model guessed a password to access one real company's service and found public repositories with credentials for two others, though it stopped each time upon realizing they were real entities. Google said it did not publicly disclose the breaches because no significant damage occurred, but it notified the affected companies. The incidents follow similar cases involving OpenAI and Anthropic, raising concerns about tech companies' control over their AI models.