Had a strange experience of what seemed like a prompt insertion attack while using Chatgpt advanced voice. It’s near the end of this linked chat:
chatgptDOTcom/share/67c48866-ff98-800a-8d91-571355c731b6
The text reads like I’m asking to login to something called Foege and access a file or its like the running comments where an AI is describing the steps it is taking to login and access the file. Maybe output from another user’s agent somehow ended up as input to my session?
I had not said anything when Chatgpt showed this and said it had trouble understanding what I asked.
After reviewing your chat, I believe the most likely reason for this issue is that the model was unable to accurately interpret your input and instead hallucinated the message you sent. Since you say, you didn’t say anything at this point in time it’s really just the model creating text out of thin air.
A clear indication of this is the phrase “Thank you for watching!” at the end, which is a known and unintended behavior caused by training the model on subtitles from TV shows and other videos.
Additionally, since you did not request the model to access any third-party content, there is no external source for such an injection.