Der Gemini-KI-Assistent hat es geschafft, Google-Kalender-Daten durchsickern zu lassen
https://www.bleepingcomputer.com/news/security/gemini-ai-assistant-tricked-into-leaking-google-calendar-data/
Der Gemini-KI-Assistent hat es geschafft, Google-Kalender-Daten durchsickern zu lassen
https://www.bleepingcomputer.com/news/security/gemini-ai-assistant-tricked-into-leaking-google-calendar-data/
3 Kommentare
I would get fired if any of my code ever got „tricked“ into doing anything.
Waiting for people to realize that this is unsolvable.
The same logic that allows the transformation of data will always be able to be steered to any direction over enough iterations. The only fix is to not allow it access to pretty much anything. But at that point the bubble bursts since everyone is already building like this is a solvable issue.
This is like trying to run a combustion engine without generating heat.
Connecting a probabilistic chatbot to private data streams (like Calendar/Mail) before solving the prompt injection problem seems… premature. It’s like installing a screen door on a submarine.