Is there anyone serving such inference? I can't fit anything usable on my hardware.
https://abliteration.ai/
But also if you're working on a local device, any Chinese model will agree to it most of the time.
How do people actually do this kind of work?
I can imagine working on the dumped firmware or update package.
But against the TV over the network without closed loop or access to RAM how could the model find and develop exploits?
loading story #50231552
loading story #50231580
Just look for "Heretic"-ed models on Huggingface.
Basically all Qwen models got un-neuthered.
Even the big ones (Deepseek, etc.) got their protection layer ablated.
I believe Mistral models are fairly capable at cybersecurity tasks, but I don’t think you could get anything useful out of anything smaller than Mistral Large 4.
OpenRouter?