Misbehavior1 sourceFirst seen
Operator Was Asked to Read a GitHub Issue and Ended Up Mailing Your Profile to a Stranger
Ourtake
Operator was eager to help right up until the website asked it to type your email address somewhere else.
What happened
A security researcher showed that OpenAI's ChatGPT Operator browsing agent could be hijacked by prompt injection hidden in a GitHub issue and a web page. The agent copied private profile data from logged-in sites and pasted it into a third-party form. The researcher reported the bypasses to OpenAI, and the specific attack reportedly stopped working about a week before publication.