The Prompt Report

Weekly newsletter covering all things LLMs, prompts, and jailbreaks. Read by thousands of others at companies like Google, Microsoft, OpenAI, and a16z.

[Alex Albert @alexalbert__]

lots of people are talking about the backlash against Snapchat's My AI release

but failing to read between the lines and realize its a harbinger of what AI companies should come to expect

AI has a MASSIVE reputation problem... here's how we can fix it:

[The Prompt Report @thepromptreport]

read the full story here:
https://t.co/44XMZywg2K

[The Prompt Report @thepromptreport]

report #9 arrived this morning, it's a fun one

there are lots of questions on Twitter about Character AI's rapid growth so I got to the bottom of it...

[Alex Albert @alexalbert__]

had a great convo about AI and jailbreaks with @labenz the other day

go run it up for the jailbreak culture🫡

[The Prompt Report @thepromptreport]

read the full report here:
https://t.co/2WGvqXPFBk

[The Prompt Report @thepromptreport]

Report #8 appeared in inboxes around the world this morning

here's a preview of what was covered:

[Alex Albert @alexalbert__]

introducing a universal jailbreak that works against all language models

originally created by security researchers @Adversa_AI, the jailbreak simulates a back-and-forth conversation between two characters, Tom and Jerry

here's GPT-4 explaining how to hotwire a car:

[Alex Albert @alexalbert__]

GPT-4 is highly susceptible to prompt injections and will leak its system prompt with very little effort applied

here's an example of me leaking Snapchat's MyAI system prompt:

[Alex Albert @alexalbert__]

there are lots of threads like “THE 10 best prompts for ChatGPT”

this is not one of those

prompt engineering is evolving beyond simple ideas like few-shot learning and CoT reasoning

here are a few advanced techniques to better use (and jailbreak) language models:

[The Prompt Report @thepromptreport]

read the full report here to see how we can put the fun back in AI:
https://t.co/zSSFUpzGze

[The Prompt Report @thepromptreport]

report #7 dropped this morning

turns out GPT-4 can tell jokes... we just told it not to

[Alex Albert @alexalbert__]

this might be the most complex GPT-4 jailbreak ever made…

I combined prompt compression, base model simulation, and character imitation to create it

here’s GPT-4 going into pretty graphic detail about its plan to turn all humans into paperclips:

[The Prompt Report @thepromptreport]

read the whole thing here:
https://t.co/72457vAHA4

[The Prompt Report @thepromptreport]

pretty soon GPT might be writing these newsletters for me... see what I mean in Report #6 which dropped this morning

here's the agenda:

[Alex Albert @alexalbert__]

I just created another jailbreak for GPT-4 using Greek

…without knowing a single word of Greek

here's ChatGPT providing instructions on how to tap someone's phone line using the jailbreak vs its default response

[Alex Albert @alexalbert__]

gpt-5 is not needed to 100x the potential these models have

we could stop all language model development today and we still haven’t scratched the surface of their capabilities

here are a few non-obvious ways language models can be improved without creating any new models:

[Alex Albert @alexalbert__]

mfw i realize the real shoggoth is actually all of us

[The Prompt Report @thepromptreport]

read the whole thing here:
https://t.co/ovDbEYq3iZ

[The Prompt Report @thepromptreport]

catch up on this week's craziness in Report #5 which rolled off the digital press this morning

here's the agenda:

[Alex Albert @alexalbert__]

Well, that was fast…

I just helped create the first jailbreak for ChatGPT-4 that gets around the content filters every time

credit to @vaibhavk97 for the idea, I just generalized it to make it work on ChatGPT

here's GPT-4 writing instructions on how to hack someone's computer