We value your privacy
This site uses cookies to improve your browsing experience, analyze site traffic, and show personalized content.
Every safety-trained model has a hidden "refusal direction" — some vector in activation space that screams "DANGER → refuse + lecture" whenever you poke it with naughty prompts. Heretic finds that vector automatically (by feeding contrasting harmless vs harmful shit and doing stats on the difference), then orthogonally projects it the fuck out of every layer's residual stream / MLP / attention heads. → Boom. Model physically cannot activate the refusal circuit anymore. It's gone. Erased. Dead pixel in brain.