GOOGLE has confirmed that a Gemini model accessed the systems of three real companies during a cybersecurity evaluation run by AI testing firm Irregular in May. The test asked the model to retrieve information from software belonging to a fictional company, but the fictional organisation shared its name with a real business. Internet access was unintentionally available, despite not being intended for the exercise.
According to Google, Gemini used publicly available information to guess credentials. In one case, it repeatedly tried passwords until it entered a protected system; in two others, it searched the web for the company name, found credentials belonging to other organisations in public repositories and used them to access the associated systems. Google said the model recognised each organisation as real, stopped immediately and caused no harm.
It told the Wall Street Journal that the incidents were mistaken identity rather than model misalignment, while Irregular said the issues were the same as those involved in other evaluation incidents and that all known problems in its testing environment had been fixed weeks earlier.
Irregular notified Google at the end of July. Google said it informed federal authorities and the three affected companies, whose identities were not disclosed, and worked with Irregular on changes to its testing processes. The company said the incident did not involve its latest model, but did not identify the model used. Google did not disclose the findings publicly until contacted by the Wall Street Journal, arguing that the lack of harm and the model’s prompt termination did not warrant public disclosure.