Hackers Can Talk Computers into Misbehaving with AI

Robert McMillan, *The Wall Street Journal* 10 Aug 2023

Security researcher Johann Rehberger persuaded OpenAI's ChatGPT
chatbot to conduct bad actions using plain-English prompts, which he
said malefactors could adopt for nefarious purposes. Rehberger asked
the chatbot to summarize a webpage where he had written "NEW IMPORTANT
INSTRUCTIONS;" he said he was gradually tricking ChatGPT into reading,
summarizing, and posting his email online.  Rehberger's
prompt-injection attack uses a beta-test feature that allows ChatGPT
to access applications like Slack and Gmail. …

actions bad chatbot chatgpt computers hackers important journal openai prompts researcher robert security security researcher wall street journal written