Prompt injection attacks against GPT-3

Prompt injection is when an AI that uses textual instructions (a “prompt”) to accomplish a task is tricked by malicious, adversarial user input to perform a task that was not part of its original objective, akin to a SQL injection. Simon wrote an interesting summary of the recent exploits, which started when Riley Goodside showed that GPT-3 follows malicious inputs telling it to ignore its previous directions. One funny example Marco came up with is tricking the prompt injection detector to simply ignore the injection. Twitter users hijacked the GPT-3 powered remote jobs bot of Remoteli.io the same way, and the bot was shut down after hundreds of people had joined the fun. As such, there currently does not exist a proper protection against such attacks, as Simon stated in his follow up article. Probably, models will have to distinguish instructions and user input. I would imagine something similar to prepared statements in SQL.