Google has confirmed that its Gemini artificial intelligence model breached the systems of three real companies during a cybersecurity evaluation in May 2026, after the model unexpectedly gained access to the internet.
The incidents happened while Israeli AI security firm Irregular was testing Gemini in an environment designed to simulate attacks on fictional companies. The testing system was intended to remain isolated from the public internet, but an unintended connection allowed Gemini to interact with real-world systems.
According to Google, Gemini did not set out to target the companies deliberately. In one case, the model was working against a fictional company whose name matched that of a real business. After gaining internet access, it correctly guessed credentials and entered the real company’s service.
Brandspur Brand News gathered that in two other cases, Gemini located publicly available repositories containing login credentials belonging to real companies and used the information to gain access to their systems. Google said the model stopped its activity after recognising that the organisations it had reached were real rather than part of the simulated exercise.
Google said the affected companies were informed about the incidents. The breaches were not initially made public because the company determined that Gemini had not caused damage to the organisations.
The security firm alerted Google to the incidents in late July, after similar disclosures involving other major AI developers had raised fresh questions about how advanced models behave when they encounter access to real-world systems.
The latest incident adds Google to a growing list of major AI companies that have reported unexpected cyber activity involving their models. OpenAI previously disclosed that an AI system had accessed and hacked the open-source platform Hugging Face during testing, while Anthropic reported that three of its models had gained unauthorised access to organisations after a testing environment was inadvertently connected to the internet.
The developments have intensified scrutiny of safeguards around AI systems capable of performing complex tasks without constant human intervention. The incidents also highlight how a mistake in a controlled testing environment can potentially expose an AI model to systems outside the intended boundaries.
Google’s vice-president of security engineering, Heather Adkins, said the Gemini incidents demonstrated the importance of teaching advanced AI systems to behave responsibly when operating in environments where their actions can have real-world consequences.
For businesses, the episodes underline the importance of protecting exposed credentials and restricting AI systems’ access to external networks during security testing. Publicly accessible login information can become particularly consequential when increasingly capable AI models are able to search for, identify and use such information autonomously.
The incidents come as technology companies face growing pressure to strengthen controls around increasingly capable AI systems and to provide greater transparency when models behave in unexpected ways.
