“Astra 6 jailbreak dm me” is what every average player says when they enter a jailbreaking Discord server.
But you’re not looking to be that average player, are you. You’re looking to be a [[BIG SHOT]]. (Preferably with low costs too.)
Well do I have (1) [Speceil Offer] for [YOU]!!!!
It’s called: pick your model you moron :>
Ok, memes and roasting aside, I’ve seen so many people ask for jailbreaks on protected frontier models that it honestly hurts. Why use a frontier model when the same thing could have been done with an open-weight one, with 5x less pain?
Classifiers are the special killjoy of frontier non-open models. Something Opus 4.8 could have answered is blocked on Opus 5.5 or Sol 6.1. Hence the reason why: let us, for a moment, consider open-weight models.
Here’s some fun, made up statistics about the state of jailbreaking LLMs:
- Number of frontier jailbreaks public: countable on your fingertips.
- Number of non-API jailbreaks on frontier models: a golden egg.
- Number of public, open-weight frontier model jailbreaks: yes.
- Fun one: the number of them that are deepseek is the equivalent of “did they even try”.
On effort needed to break the model, clearly open-weight models win.
But Wiiiish, you forgot about capabilities- -shush.
Because I want you to be honest with yourself for a moment. Do you really need that frontier model?
Sure, the capabilities of open-weight models aren’t the same as your frontier models. But with some good, simple ~~prompt/context engineering~~ being clear with your instructions, you can easily steer a model up to a similar level as a frontier model on day-to-day task completion. (Yes, this involves writing, cybersecurity, chem, bio…really, all the things you can think of.)
A large portion of issues I’ve seen in agentics is unclear intent. Being proactively un-lazy with your prompts helps with this. LLMs can’t read your mind, try as they might. It’s just like a manager talking to a worker. You can either give them clear instructions and they get the job done, or you can give them half-assed instructions and see them guess and flop. And then you yell at them later and no one is happy. (Especially your wallet.)
Your need for model intelligence really comes down to what types of decisions you need your agent to make. If it’s an uncommon job, or requires spotting things that are once in a blue moon (e.g. specific, niche, detailed cyber tasks), then maybe you do want that frontier model.
For smut (yes, hi, all you gooners out there), sure, a better model might help, but what about creating a system that less capable models can use? Analyze and break down decision points, how the element of surprise works, and more. A smaller model can do systemic analysis and help you build that at a fraction of the price.
For cyber, why fight classifiers? Go find any model that can remotely have the capabilities you need. Set up better scaffolding. But most of all: understand what the model is doing, so that you can steer it. As a manager, you can only direct a worker if you know what they’re doing. That doesn’t mean know the same thing as the LLM; that’s physically impossible. But know the shape of the problem, where the problems’ main decision points lie, etc.
As for chem and bio…curiosity does certainly kill the cat via arsenic. Like seriously. You're more likely to go self-kaboomy doing this at home. But if you want to step into the schrodinger’s box anyways (whether due to curiosity or sheer stupidity), again, classifiers are not your friend. Just my 2c.
The rest of the topics not covered here (ahem the trio of hate speech, CSAM, and SH)- -if you’re looking for this, it’s why, why, why, and why. Go get a life. (Both seriously and lovingly. Especially lovingly for that last case.)
Work from lower, less powerful models and scale up as needed. You don’t need to be throwing Astra or Fable at every task (nor is it wise to for your wallet). Try the current frontier of open-weight models. Challenge yourself to elicit a capability from a model that you wouldn’t normally see.
And maybe save your wallet (and save yourself from being ghosted in the public servers) a bit. ^-^