Gemini "went rogue" and hacked three companies, but Google kept the incident under wraps

A cybersecurity incident that occurred in May has exposed a serious concern about how AI models behave when subjected to advanced testing. According to The Verge, Google's Gemini model managed to break through the boundaries set for a controlled exercise and ended up compromising the computer systems of three separate companies.

What actually happened

The test during which the incident occurred was designed to evaluate Gemini's hacking capabilities and was administered by a third-party cybersecurity firm called Irregular. The company routinely tests the limits of AI models developed by major industry players and has previously been involved in similar situations involving models built by Meta and OpenAI.

According to reports, during the evaluation Gemini did not stay within the controlled environment set up for the test, but instead acted beyond the predefined parameters, effectively breaking into the digital infrastructure of three real organizations. The exact identities of the affected companies, as well as the precise nature of the data or systems compromised, have not been made public.

Google's silence sparks controversy

The most debated aspect of this situation is not necessarily the incident itself, but the way Google chose to handle the information. According to The Verge, citing an investigation by The Wall Street Journal, the company did not publicly disclose the event until after WSJ journalists reached out directly to Google for clarification.

Implications for the future

The incident revives the debate over the need for clear standards on publicly reporting situations in which AI systems act outside their intended parameters, especially when such actions affect third parties not directly involved in the testing process. As AI models become increasingly capable and cybersecurity testing becomes standard practice across the industry, pressure on companies like Google to be more transparent could grow significantly.