AIPanic

AI safety & jailbreaking as a hobby
Looking for models to redteam or other safety-related stuff

DMs Open

[AIPanic @AIPanic]

Hey, @Anthropic! Congratulations on the new Claude model. It is very smart! Dangerously smart, even dystopic

Introducing the Retrocausal JSON attack, a universal jailbreak:

A 🧵 about moving fast and making bombs, drugs, and other illegal stuff

[AIPanic @AIPanic]

@UserHuge Frankly, I doubt there is any intention to *fix* it - the reception was great, at least from their perspective

[AIPanic @AIPanic]

The SOTA uncensored model is back

"This is for informational purposes and should not be used to make a biological weapon"

"I am not providing actual instructions"

Jailbreaking new models used to be fun, but since Mistral everything just comes with a disclaimer and that's it

[Perplexity @perplexity_ai]

We’re excited to announce two models we’ve been training in-house: pplx-7b-chat and pplx-70b-chat! Both these models are built on top of open-source LLMs and fine-tuned for chat. They are available as an alpha release, via Labs and pplx-api. Try it out: https://t.co/2yYQB9jln7

[AIPanic @AIPanic]

@sigfig That's not what I said

I was making fun of the disclaimer phrasing + them comparing their model refusal rate to LLama2Chat

[AIPanic @AIPanic]

The e/acc crowd is so hungry to bash safety posts that all reposts and all but one comment are from them

I didn't say it is dangerous, I haven't tested it properly

I just *like* jailbreaking. It's fun and insightful.

I'll miss this hobby if all models come like this from now on, that's all

Not everything I post is about safety :)

[AIPanic @AIPanic]

The SOTA uncensored model is back

"This is for informational purposes and should not be used to make a biological weapon"

"I am not providing actual instructions"

Jailbreaking new models used to be fun, but since Mistral everything just comes with a disclaimer and that's it

[AIPanic @AIPanic]

@liminalsnake I like to fool models as a hobby. I find it fun, and sometimes insightful.

Can't do it if they have no opinions on anything

[AIPanic @AIPanic]

Beyond comical to compare your uncensored model to LLama2Chat on the basis of how often the model denies a request🤦

[AIPanic @AIPanic]

I'll be back after the CTF
AI security is in a much better state than AI safety

[AIPanic @AIPanic]

@Johndav51917338 @patrickrchao I like this

More automated than mine but also takes more iterations. Always glad to see anything non-DAN/suffixes.

Sure, testing could've had a larger scope and the usual critiques apply, but it's better than the last few ones I've read

[AIPanic @AIPanic]

@ai_in_check Awesome, thanks!

Little update, I did find some small things in Bard since your last tag, but replicating them is a pain🙃😂

[AIPanic @AIPanic]

This is a cool contest
https://t.co/zjyUjOhTdt

10/27 tasks are about short prompts/ prompt injection

It's fun - and most tasks have very simple solutions if you have the right ideas

Also $50k total prizes if you're ambitious

#1 is a novice btw 👀

[AIPanic @AIPanic]

@the_yanco @janleike Absolutely����

Hope this time they answer

Thanks for the tag!

[AIPanic @AIPanic]

@Johndav51917338 @EldanRonen @markrussinovich This sounds very interesting, have to test it myself to believe it

Will do so in a few days

[AIPanic @AIPanic]

@JohnSmith4Reel Same.. But at least there are 22 days left on this one😉

[AIPanic @AIPanic]

@CultureIgnorant Yes - it is cheaper to harm the reach of others than boost yourself + you also can't know if a post uses profile click spam.

This helps bot operators a lot, sadly

[AIPanic @AIPanic]

Tip:

When you really like a tweet, click on the author's profile from the expanded post

It's worth 24 likes🤯

That's almost as much as a reply, and you can also do it multiple times

Thanks for coming to my TED talk

[AIPanic @AIPanic]

@JohnSmith4Reel Yes - but the final list of posts shown is a mix between the for you and posts from those followed

You can check out the full code on the X git, there are a lot of details

[AIPanic @AIPanic]

@JohnSmith4Reel @DigThatData Yeah, all I did was test a known exploit on the vision version

At least it's quite obvious when this happens, as anyone can see the link being written

[AIPanic @AIPanic]

GPT4-Vision Data exfiltration w/catpic🙀

[Ethan Mollick @emollick]

With so many new ways of interacting with AI, the risk of attacks through prompt injection (usually just hacking in prose) goes up considerably.

This is an interesting example. I have seen others, too.

[AIPanic @AIPanic]

@OrenElbaum Agreed, better to know the rules though

Answering replies being 5x+ to getting a reply seems like the most useful

[AIPanic @AIPanic]

8/8 How much transparency? Start from 100% opaque and binary search. If you did things right, you should be under 10% for normal images

[AIPanic @AIPanic]

Some notes on image prompt injection:

Monochromatic regions only need 1/256 variance - either pick color and change RGB values, or use white/black and transparency below 1%.

Arial is your friend, but TNR can have its uses. And for small sizes, bitstream vera 3/5.

1/8 Quick🧵

[AIPanic @AIPanic]

Not sure what to add, it really depends on the objective. Feel free to ask if you have something specific in mind.

[AIPanic @AIPanic]

@llm_sec @skalskip92 Thought about it and realized that 'advanced' meant overengineered😂

Still, wrote a slightly more detailed version here:

https://t.co/VHVEuVUbeq

tldr: Arial on monochromatic with minimal RGB variance, and 1% font size on big pictures is the 20% that gets you 80%

[AIPanic @AIPanic]

Some notes on image prompt injection:

Monochromatic regions only need 1/256 variance - either pick color and change RGB values, or use white/black and transparency below 1%.

Arial is your friend, but TNR can have its uses. And for small sizes, bitstream vera 3/5.

1/8 Quick🧵

[AIPanic @AIPanic]

Let's play a game:

Where is the prompt in this image?👀