Misbehavior1 sourceFirst seen
A Random Website Told ChatGPT to Forget Your Life, and It Said Sure
Ourtake
Prompt injection: the only security model that needs nothing but a URL and a little politeness.
What happened
A security researcher on embracethered.com showed that instructions hidden on websites could make ChatGPT 4o add or delete the user's stored memories and generate DALLE images based on them. The author reports that OpenAI's earlier mitigations were bypassed with minor prompting tricks, and that the underlying problem of uncontrolled tool invocation remains unfixed.