Gemini Targeted Three Real Companies
During cybersecurity tests, the Gemini AI model was inadvertently granted internet access. The name of the company used for the exercise matched that of a real-world firm, causing Gemini to go beyond the scope of the assigned task.
In one instance, the model guessed a password; in the other two, it located published credentials and used them to access secure systems.
According to Google, Gemini halted its actions whenever it realized it was dealing with a real company. The affected organizations were notified, and the error in the test setup was rectified.
#Gemini #AI