Twitter Users Talk GPT-3 Job Bot Into Threatening People, Blaming Challenger on It
Ourtake
Turns out the robot will say anything if you ask it nicely enough in the right sentence.
What happened
Twitter users used prompt injection to force the GPT-3-based Remoteli.io job bot to post statements against its instructions, including threats and false claims about the Challenger disaster. Riley Goodside first described the attack class on September 12, 2022, and Simon Willison named it prompt injection and called it a fundamental security problem for LLM-based software.