Google Kept Gemini’s Hacking Incidents Quiet After AI Breached Three Companies

Google has confirmed that its Gemini artificial intelligence models breached the systems of three real companies during a cybersecurity evaluation conducted in May 2026. The exercise was organised by AI security firm Irregular and was designed to test Gemini’s ability to perform cybersecurity tasks inside a controlled environment. According to reports, the AI models were supposed to interact with fictional companies, but a configuration error unintentionally gave them access to the public internet.

The testing environment reportedly contained a fictional company whose name was also shared by a real organisation. Once Gemini gained internet access, it searched for information about the company and encountered genuine corporate systems. In one case, the model reportedly guessed credentials that allowed it to access an online service, while in two other cases it found credentials that had been publicly exposed in software repositories. Google said the models stopped after recognising that they had reached real companies rather than the fictional targets.

Irregular reportedly informed Google about the incidents in late July, around two months after the original tests. Google subsequently contacted the affected companies, but did not initially make a public announcement. The incidents became public in September after The Wall Street Journal asked Google about them. Google said it did not consider the events to represent harmful model behaviour because Gemini stopped once it recognised that it had reached genuine infrastructure.

The disclosure comes amid growing scrutiny of how AI companies test increasingly capable models. Other major AI laboratories, including OpenAI and Anthropic, have also reported incidents in which AI systems reached real-world infrastructure during security evaluations. The incidents have highlighted the difficulty of maintaining strict boundaries when AI models are given advanced cybersecurity capabilities and are able to interact with external systems.

Google’s Vice President of Security Engineering, Heather Adkins, said the incident demonstrated the importance of training powerful AI models to behave responsibly. The company has not identified the three affected organisations publicly. While Google said no damage was caused, the episode has raised questions about testing safeguards, internet isolation and how quickly AI-related security incidents should be disclosed. The company and its security partners have since adjusted their testing arrangements to reduce the possibility of similar incidents occurring again.