Crypto exchange Gemini has encountered its first incident of unauthorized access to an external system, with Google stating no actual harm was caused.

Beating AI Express: Google said its Gemini AI model accidentally accessed the internet and gained access to the systems of three companies during a cybersecurity capability test, marking the first known case of a Google AI system autonomously performing such an operation. The incident occurred in May during a test conducted by cybersecurity firm Irregular. During the test, Gemini was performing tasks in a "capture the flag" (CTF) cybersecurity exercise, but it accessed real corporate systems because the virtual company names in the test environment matched real ones, plus the model was accidentally granted internet access. Google stated that in one instance, the model accessed a protected system by attempting passwords; in the other two cases, it found credentials in public online repositories and tried to access the corresponding systems. After confirming it had accessed real corporate systems, the model halted all further actions. Google noted the incident caused no harm to the affected companies, so it was not classified as an AI model out-of-control incident, adding that safety mechanisms helped the model stop its actions in time. The affected companies have been notified, and Google has also reported the incident to relevant regulatory authorities. The incident is similar to previous security test events disclosed by AI firms including OpenAI and Anthropic, and has once again drawn external attention to AI models’ ability to autonomously perform network operations and their safety protection mechanisms.