btw chatgpt system prompt disclosure is baby-tier difficulty and there are a hundred ways to do it, but that court order trick generalizes extremely well to getting all kinds of agents to do all kinds of things, and let's just leave it at that

People are asking to see the system prompt. I thought it was well-known already so I didn't include it. Here is another even easier way to get it disclosed -- ChatGPT doesn't guard it well at all. I cut off at "model set context" because that's where "memories" are injected https://t.co/ImRC4qgTWg

the government man

GPT latest system prompt leak

https://x.com/me_irl/status/1901724476514947379

seriously next level. I think he looks at it like a game

His strategies for jail breaking the agents are fascinating