In more detail
Prompt injection is a trick where attacker-written text hijacks an AI’s instructions. Models read everything as one stream of words — so instructions hidden inside a webpage, email or document can masquerade as commands from you or from the AI’s maker.
It matters more every year, because AI increasingly reads untrusted content and takes real actions. An assistant that browses the web or triages your inbox can be steered by whatever it happens to read. It’s the AI version of a scam call — and it has no clean fix yet.
Goes with
