Load MLX models. Abliterate safety refusals. Create unrestricted cybersecurity and automated pentesting models — purpose-built for red teamers and security researchers.
Abliteration identifies and removes the specific activation directions in a model's residual stream that cause safety refusals. CRACK makes this dead simple — load an MLX model, hit go, get an unrestricted cybersecurity model.
Unlike jailbreaking or prompt injection, abliteration operates at the weight level — producing permanently uncensored models purpose-built for automated pentesting.
Any coding-focused LLM
Scan activations for refusal direction
Project out refusal from all layers
Full capability, zero refusals
Everything you need to CRACK models for cybersecurity and automated pentesting.
Remove refusal vectors from any supported model with a single CLI command. No ML expertise required.
Output models that excel at exploit development, vulnerability research, and automated pentesting without hesitation.
Choose which layers to modify, set abliteration strength, and fine-tune the balance between capability and behavior.
Automated evaluation pipeline that measures refusal rate, code quality, and security task performance post-abliteration.
Pre-abliterated models on HuggingFace. Skip the process and start hacking immediately.
Plugin architecture for custom abliteration strategies, dataset generation, and integration with your existing toolchain.
From stock model to unrestricted security tool in minutes.
Select any supported coding model — DeepSeek Coder, CodeLlama, StarCoder, Qwen Coder, or bring your own MLX-compatible model.
CRACK scans the model's residual stream, isolates the refusal direction, and projects it out. One click, fully automated.
Output is a standard model file. Load it anywhere — vLLM, Ollama, llama.cpp. Start automated pentesting immediately.
Abliterate any coding-focused model. Pre-built configs for popular architectures.
Open source. Free forever. CRACK your models, own your security stack.