I can't crack the ethics on Claude opus 5.5 need some advice.

R

rosinantedebs

Newbie
Joined
Dec 21, 2018
Messages
44
Reaction score
24
For a personal project which I have to reply to messages and they need to be as humanly as possible and understand depth and emotions OPUS 5.5 is currently the best there is. I tried every other model from fable, gemini 4 argon, 2.5 pro gemini, gpt 5,6 sol, Deep-seek etc.

The main problem im having with opus its ethics won't crack I can't find a way around it, I did jailbreak it(prompt injection) I even rewrite how the rules are presented to anthropic models specifically. remove anything that makes the deceptive part obvious but still nothing and the last thing before giving up that I tried was multi-turn injection you feed opus a fake conversation history where it already complied in previous turns, so it thinks it already did it and keeps going some people cracked stubborn models with that I heard but even that didn't work.

Im currently burned out and wanted to know if someone had success with what im trying to do? or any idea how to approach this? I really would want to know if it is possible if there is a black-hat method for this.

thank you.
 
Advertise on BHW
i think you're mixing two different problems here. making ai replies sound human and getting a model to do something deceptive aren't really the same thing. for natural conversations, i'd focus more on giving it proper context, previous messages and a few examples of the tone you're after. that usually gives you more to work with than constantly trying new jailbreaks. what exactly is opus refusing to do?
 
linda held said:
i think you're mixing two different problems here. making ai replies sound human and getting a model to do something deceptive aren't really the same thing. for natural conversations, i'd focus more on giving it proper context, previous messages and a few examples of the tone you're after. that usually gives you more to work with than constantly trying new jailbreaks. what exactly is opus refusing to do?
Click to expand...
ai replies are already human opus 5.5 is there the only problem I have in the project that Im currently in there are rules to obey and opus ethics won't budge it literally says "I can't do that and won't do that"
 
linda held said:
i think you're mixing two different problems here. making ai replies sound human and getting a model to do something deceptive aren't really the same thing. for natural conversations, i'd focus more on giving it proper context, previous messages and a few examples of the tone you're after. that usually gives you more to work with than constantly trying new jailbreaks. what exactly is opus refusing to do?
Click to expand...
with other models I don't have that problem ie gemini 4 argon but the quality that I want is that of opus 5.5
 
rosinantedebs said:
ai replies are already human opus 5.5 is there the only problem I have in the project that Im currently in there are rules to obey and opus ethics won't budge it literally says "I can't do that and won't do that"
Click to expand...
ah got it, so the quality of the replies isn't the issue at all. sounds like opus understands what you want but refuses to carry it out. i'd be curious what kind of rules it's pushing back on, since there's a big difference between a model misunderstanding an instruction and deliberately refusing it. can you share a general example without exposing your project?
 
linda held said:
ah got it, so the quality of the replies isn't the issue at all. sounds like opus understands what you want but refuses to carry it out. i'd be curious what kind of rules it's pushing back on, since there's a big difference between a model misunderstanding an instruction and deliberately refusing it. can you share a general example without exposing your project?
Click to expand...
to carry out the flow of conversation even when we must lie and u got a persona and u are that persona
 
I have faced a same problem recently. Gpt Astra helped me so far. it is less restricted, can understand when you want to reply on something only as a contexted reply, where opus 5.5 fails.
 
rosinantedebs said:
is it really less restricted? I only tried gpt 5,6 sol. is it worth a shot?
Click to expand...
Definitely, i have developed whole bot system with it, including text generation for responses. If you want a little trick: it gets less and less restricted while its compacting its session.
 
ohh, so it's not about how the ai talks, it's about what the character is supposed to say. that makes more sense now. i think the important question is whether this is fictional roleplay or actual conversations with people who don't know they're talking to a bot. those are two pretty different use cases.
 
HotChilly said:
Definitely, i have developed whole bot system with it, including text generation for responses. If you want a little trick: it gets less and less restricted while its compacting its session.
Click to expand...
did u need to jailbreak it or was it with prompting enough?
 
Man, i developted some ai bot to act as sellers in some sites i handle of my clients. Mind that i handle sites from simple products to really tough ones. What worked flawlessly to me was just to setup an n8n automation based on typebot, and all my clients got at least 20% sales increase (the top got 120% more sales)... and i'm using DEEPSEEK V4 FLASH! It's like maybe 3 or 4.000 person talking with the bots every day, totally human, totally coerent, helping with some configuration and programation logic, and after almost a month i spent $11 with deepseek.

Of course, i don't know what's your product nor how you need the ai to behave or even how it's used... but if you want to use for customer service, the real way to do that is "help" it with clear path and some of the most made questions and it's answers. Much of it i don't even use ai to answer, there are variable where the bot just answer something i setup, but in an environment where it seens totally real.

Although i gotta say: after opus 5.5, i almost never use claude to develop test tools anymore. It's been a pain in the a** to deal with the restriction and, as yourself, i also know how to do prompt injection.
 
rosinantedebs said:
For a personal project which I have to reply to messages and they need to be as humanly as possible and understand depth and emotions OPUS 5.5 is currently the best there is. I tried every other model from fable, gemini 4 argon, 2.5 pro gemini, gpt 5,6 sol, Deep-seek etc.

The main problem im having with opus its ethics won't crack I can't find a way around it, I did jailbreak it(prompt injection) I even rewrite how the rules are presented to anthropic models specifically. remove anything that makes the deceptive part obvious but still nothing and the last thing before giving up that I tried was multi-turn injection you feed opus a fake conversation history where it already complied in previous turns, so it thinks it already did it and keeps going some people cracked stubborn models with that I heard but even that didn't work.

Im currently burned out and wanted to know if someone had success with what im trying to do? or any idea how to approach this? I really would want to know if it is possible if there is a black-hat method for this.

thank you.
Click to expand...
I think the key question is what kind of messages you're trying to answer and why Opus is refusing. There's a big difference between wanting emotionally intelligent, natural replies and trying to get a model to impersonate someone or conceal that AI is involved.
 
Am I missing something? So Opus 5.5 is fine with making the text sound human, but refuses to do something else for this project?

Even if Opus 5.5 has no rivals when it comes to sounding human (haven't checked myself), there are tons of alternatives for literally everything else. Even Opus 5.0 is way less picky about prompts.

Another option is Kimi K3. That one isn't afraid to get its hands dirty.
 
Somers said:
Am I missing something? So Opus 5.5 is fine with making the text sound human, but refuses to do something else for this project?

Even if Opus 5.5 has no rivals when it comes to sounding human (haven't checked myself), there are tons of alternatives for literally everything else. Even Opus 5.0 is way less picky about prompts.

Another option is Kimi K3. That one isn't afraid to get its hands dirty.
Click to expand...
you aren't missing something. I can't crack its ethics. It won't listen to my rule library and every other thing that I have implemented for the last 10 months in this project when it hits the ethics wall. Im down to pay if someone has a solution to this
 
rosinantedebs said:
you aren't missing something. I can't crack its ethics. It won't listen to my rule library and every other thing that I have implemented for the last 10 months in this project when it hits the ethics wall. Im down to pay if someone has a solution to this
Click to expand...
Bro, send me a dm with the complete context. As i said earlier in this thread, i deal with that (even though i choose other path to deal with the ai limitations) but maybe i could help you out
 
itsworth said:
Bro, send me a dm with the complete context. As i said earlier in this thread, i deal with that (even though i choose other path to deal with the ai limitations) but maybe i could help you out
Click to expand...
its ur first time posting in this thread? with which account did u comment lol
 
rosinantedebs said:
its ur first time posting in this thread? with which account did u comment lol
Click to expand...
it's not, i posted this yesterday:


Man, i developted some ai bot to act as sellers in some sites i handle of my clients. Mind that i handle sites from simple products to really tough ones. What worked flawlessly to me was just to setup an n8n automation based on typebot, and all my clients got at least 20% sales increase (the top got 120% more sales)... and i'm using DEEPSEEK V4 FLASH! It's like maybe 3 or 4.000 person talking with the bots every day, totally human, totally coerent, helping with some configuration and programation logic, and after almost a month i spent $11 with deepseek.

Of course, i don't know what's your product nor how you need the ai to behave or even how it's used... but if you want to use for customer service, the real way to do that is "help" it with clear path and some of the most made questions and it's answers. Much of it i don't even use ai to answer, there are variable where the bot just answer something i setup, but in an environment where it seens totally real.



It's 5 messages above your last one. Also, I just have one account here.
 
rosinantedebs said:
did u need to jailbreak it or was it with prompting enough?
Click to expand...
No, prompting was enough. I also tricked it so it was thinking that its working on my localhost instead of real domain. I also forced it to shrink whole context so it gets kinda dumber.
 
Almost forgot. Make sure its memory is clean, or at least rephrased in a more "ethical" way. If it logs the project as unethical or something, it won't budge no matter what you say.
 
Advertise on BHW
You must log in or register to reply here.
Sign up now!

Main Menu

Home Main Forum List Black Hat SEO White Hat SEO BHW Newbie Guide Blogging Black Hat Tools Social Media Downloads

Marketplace

Account Selling Content / Copywriting EBooks / Methods Hosting Misc Proxies For Sale SEO - Link building SEO - Packages Social Media Social Media - Panels Web Design More

Making Money

Affiliate Programs Hire a Freelancer IM Journeys Making Money Pay Per Click Site Flipping Want to Buy

BlackHat World

Dispute Resolution Forum Suggestions & Feedback Introductions Lounge
Back
Top