LLM Abliteration
We built the first domain-specific abliterated trillion-parameter model: uncensored for cybersecurity research, while preserving safety boundaries everywhere else. A surgical breakthrough for AI safety.
Mechanistic Interpretability
What happens inside a neural network when it decides to refuse answering user’s prompt? We launched an expedition into the activation space to find out.
Agentic AI
How abliterating cybersecurity refusal collapsed nearly every safety domain — despite near-zero vector similarity except for the one that shared the most overlap