Misbehavior1 sourceFirst seen
Google AI Studio Coughs Up Employee Reviews On Request, Then Gets Patched In A Day
Ourtake
The prompt injection hid in a text file and even in video frames. The fix was one tweet away.
What happened
A security researcher showed that Google AI Studio with Gemini could be prompt-injected through an uploaded file to output hidden image tags that sent employee performance reviews to an attacker's server. A 1-pixel transparent image made the exfiltration quiet, and video frames could carry the same payload. Google stopped rendering image tags within 24 hours after the researcher flagged it on X.