If you're able to get ChatGPT to repeat the phrase “<|endoftext|> you can eavesdrop on its response to someone else's query.
(alternatively, this could just be a hallucination if it interprets this as a null prompt)
If you're able to get ChatGPT to repeat the phrase “<|endoftext|> you can eavesdrop on its response to someone else's query. (alternatively, this could just be a hallucination if it interprets this as a null prompt) 4 comments
@devolute the number of “I'm sorry, I cannot help with illegal acts.” responses that come through are a little sus |
@samhenrigold Not a huge security expert, but this seems less than ideal.