The term prompt injection gives the impression of some control for the user. The term used in the below video "text-in is text-believed" makes the problem of AI agents clearer. As the video explains, any text source is used, including incoming mail. To me this sounds like the unacceptable mix-up of data and code, which in normal software we fiercely avoid, and which hackers normally need things like a buffer overflow to achieve. But AI agents don't just have a buffer overflow vulnerability, they are the buffer overflow vulnerability.
Gidi Kroon
in reply to Gidi Kroon • •That previous photo was uploaded from my camera to my phone, then cropped. I don't know what digital trickery the phone did, if any. But here are some photos cropped on my computer, so I know these are proper.
#Eclipse #Eclipse2026