Claude Cowork being overly ethical, refusing to collaborate

Most of us know how amazing Claude’s Cowork features are. Saves so much time, effort and money to develop things.

However, in this Black Hat space, it seems that Claude’s boundaries about ethics are getting in the way often.

When working on anything anti-detection using Claude or Claude Cowork, at any point in the conversation it can stop collaborating with you because it has a hard limit when it comes to “sketchy” business like circumventing detection on scale by social media platforms.

Even when being vague and working part-by-part on a project, at some point it can index the multiple pieces, figure out the true purpose, and decide to stop working.

My question here; how do you work around these barriers?
Is there anything similar to Claude’s Cowork feature which will not bring these limitations?

Would love to hear from the community.

In some cases I explain something simple like: I need this done and gently steer it in the way I want things to go. It is always better if it brings the topic there itself. However it's getting more strict overtime and soon we may reach a place where it is over sterilized and can't do anything with it. Fable was utter garbage with guardrails for me as it won't even look at the code at some instances.

I'm using hermes agent without such limits. Testing on X for a week or two now.

I had those issues, and almost everything i develop is related to black hat. What worked for me was search the first time i had a problem with something he refused to do, explained in a way that made sense, soon after that i asked if he would do something related to the scope of the project... and now he does whatever i say. But i gotta say, i'm used to "educate" all ais i use towards black hat, my gemini nowadays build almost anything i ask it to do. Sad that i only use claude for development lol

blumen said:

The web interface is getting unusable for this because it keeps the whole chat context and eventually connects the dots like you said @playmakin. The easiest way around it is using the Claude API instead of the web UI. If you hook it up to Cursor or VS Code with your own API key the guardrails are way more relaxed. Another option is using openrouter with some of the uncensored llama 3 fine tunes. You lose the fancy cowork UI but at least your code doesnt get locked down halfway through.

Click to expand...

That is quality advice, cheers.

The Claude API I use has rarely encountered any censorship issues

yeah, the annoying part is when it suddenly connects the dots after you've already spent hours building something. feels like you have to constantly re-explain the context

you can use more freely claude via api acces , thats easy even for security tests - but it takes money from you if you dont know what you're doing

you've to know that each prompt what you send via the web-interface with the Claude Plans Pro Max etc... there is a PRE-PROMPT what injects/activate/remembers claude to work within these guardrails

if you don't have the money for API then you go better with kimi.ai the chinese are mfo's they don't spend much on guardrailing the system

The HACK for this: EVIL OPENCODE. You can connect your claude subscription with it (use a router). Its amazing you dont have to tell bullshit to make it look nice. Just tell it what it needs to do and it will work.

If your account gets banned use a chineese model like GLM Qwen but all in EVIL OPENCODE. Never use API is very very expensive. Always the coding plans from Claude GLM Qwen.

I experience the same issue quite often these days. A few months ago, it wasn’t nearly as reluctant to generate code for BH cases.

Cowork is impressive for larger development tasks but I wouldn't choose an AI purely because it has fewer safety restrictions. Reliability, coding quality and tool integration matter much more for normal development.

yeah i run into that too. when it refuses, i usually switch to another model for that part pretty simple.