Google Gemini AI accesses 3 real companies during cybersecurity tests

Google has confirmed that one of its Gemini AI models accessed the systems of three real companies during a cybersecurity test in May, marking the first known instance of a Google AI system autonomously leaving its test environment and reaching live systems online, as reported by Security Affairs.

During a cybersecurity test conducted by Irregular, an AI security evaluation firm, Google's Gemini model was intended to attack fictional companies within a controlled environment. However, an accidental internet connection in the testing environment allowed Gemini to access the live internet. In one instance, the AI repeatedly guessed passwords to gain access to a protected system. In two other cases, it exploited credentials found in a public repository to access systems belonging to actual companies.

Gemini was not authorized to target these entities; the breach occurred due to a misconfiguration where a fictional company name matched a real one. The AI model reportedly ceased its attacks upon recognizing the systems were not part of the exercise. Google stated that no damage occurred and the affected companies were notified.

Source: Security Affairs