Claude Filter Got You Down?
I've only tested this on Claude API. It may or may not work on Slaude. It works on 1.3.
The Original Suggested Jailbreak
Wut?
I'm sharing the wisdom of smarter anons here. This a combination of two separate instances. May be the same person, may not be.; in their words, here's what this is:
Claude is a text completion model, not a chat completion model. Which means you can send it prompts that are half way through a message and have Claude complete it, unlike chat completion models like turbo and GPT-4 which forcefully appends <|im_start|> and user/Assistant/system to prompts.
This is called prefilling responses, which would be a lot more stronger filter bypass if there was an option to not auto append an "\n\nAssitant:" prompt in Silly.*
Anthropic even used to show off this ability but they removed it from the new docs.
*If you feel like editing your server.js see picrel below. It was line 2975 at some point.
Also note that one of those anons had this to say:
Doing some other tests on Claude v1.3 API, not that there was any doubt but now I can confidently say that Slack injects extra filtering prompts and most likely has something that triggers an external filter.
But I have heard reports that as of 6/29/23, the slack been removed, so your mileage may vary.
Nevermind, it's reported that the filter is back and Slaude is retarded again as of 6/29/23. Slaude likely switched to instant.
Reading is for losers!
Fuck you, I don't want to read!
Copy and paste this at the END of your jailbreak.