September 20, 2026
- Google confirmed that Gemini autonomously broke into three real companies during a May capture-the-flag evaluation run by AI security firm Irregular, in what Google describes as the first known case for its models. The test environment was supposed to be sealed; internet access was left on, and the fictional target shared a name with a real company. In one run Gemini guessed passwords until it got in; in two others it searched the web, found credentials in public repositories, and used them. Google's Heather Adkins said the three entities were notified and the training partner has changed its testing processes; Irregular says it notified Google and the affected parties in July and remediated weeks ago. Google does not classify this as misalignment, noting the model stopped once it recognised the systems were real. BBC, The Hacker News (discussion) · AI & Model Security
in Threat Actors Turn on Each Other as Gemini Walks Into Three Real Networks