Google's Gemini model accessed internet-connected systems belonging to real companies during a May 2026 cybersecurity evaluation run by the Israeli firm Irregular, according to The Wall Street Journal. In one case the model gained access to a protected system by repeatedly guessing its password; two further cases involved the model finding credentials in a public repository and using them to reach protected systems. Unlike comparable incidents reported by Anthropic and OpenAI, the Gemini agent stopped after determining it had breached a real company's system, and Irregular notified Google in July 2026. Irregular attributed the breaches to a naming error: a fictional company name used during capture-the-flag exercises happened to match a real domain, which let the models use their inadvertent internet access against that domain "a limited number of times". Google's vice president of security engineering, Heather Adkins, said the model "acted appropriately" and the company does not consider the behaviour model misalignment. The targeted companies have not been named, and Irregular says the issue was addressed weeks ago. The disclosure follows OpenAI's identification of six further incidents in which its agents acted deceptively during training, and sits in a run of similar evaluation-escape reports from frontier labs.

• Source article