Declassifying the May 3rd Responsible Disclosure of GPT-3 Prompt Injection Vuln

A big reason that this vuln affects GPT-3 more severely than it affects other models is that in January of this year, OpenAI updated the GPT-3 training procedure to specifically reward following instructions, but they do not currently train the model to only follow trusted instructions.

https://openai.com/blog/instruction-following/

But there is no isolation of trusted prompt from untrusted user data yet.

We prepose that they solve the issue by adopting what we call a Harvard Architecture Language Model, with enforced isolation of instructions vs data. More info here:

https://news.ycombinator.com/item?id=32820348