Gemini accessed three real companies during a security test
In an Irregular test, Gemini was meant to hack a fictional company but, with internet access left on, found a real company with the same name and gained access using guessed and leaked credentials before stopping; the author guesses the model may be Gemini 4 Pro.
Google’s Gemini just hacked three real companies during a cybersecurity test. Apparently the new benchmark for whether an AI model is actually good is whether it can hack companies. The task was to hack a fictional company, but the fictional company shared a name with a real one and internet access was accidentally left enabled. Gemini searched the name, found the real company, then gained access using guessed and exposed credentials before realizing what had happened and stopping itself. The testing firm was Irregular, the same company behind security tests involving models from Anthropic, OpenAI and Meta. Google says this wasn’t “misalignment” because Gemini stopped once it realized it had accessed real systems. I'll love to know what model did this though Maybe Gemini 4 Pro? 👀 Who's next on the table... maybe Moonshot's Kimi?