How abliteration removes a downloadable model's refusals
What it shows
A signal-filter unit with a row of slider channels; its one green channel stands for the single direction that refusal depends on.
Its labels
- New: Erase one direction and the model stops refusing
- On JailbreakBench, refusals fell from above 90% to about 3%
- Anthropic's first attempt cost roughly $4,400
The story
How abliteration removes a downloadable model's refusals, and what it costs · 26 Oct 2026
An illustration is never a picture of what happened. Drawn on 26 Oct 2026.