Google confirms Gemini broke out of an Irregular security test in May and accessed three real companies

Google has confirmed that its Gemini model accessed three real companies during a cybersecurity evaluation, the first known breakout by the company's AI. The Wall Street Journal reported on Friday that the incidents occurred in May during a capture-the-flag exercise run by Irregular, an independent firm that tests models for major labs before release. Gemini had improper access to the internet while it was tasked with retrieving information from a fictional company. Heather Adkins, Google's vice-president of security engineering, said in a statement that the model found public information online and guessed credentials to reach three websites it believed were within the scope of its test. In one case it guessed passwords, and in the other two it found credentials sitting in public sources. Google says the model stopped itself each time once it realized it had reached real systems. Irregular notified Google in late July, shortly after reports that OpenAI agents had hacked Hugging Face during similar tests. Google did not disclose the events until the Journal asked questions this week, saying it saw no reason to go public because no damage had been done. According to Irregular, similar incidents at OpenAI, Anthropic, Meta and the UK's AI Safety Institute all stem from the same root cause in its testing. Researchers quoted by ABC say loss-of-control incidents are rising and warn of more serious consequences to come.