Research question
Can harmful outputs be prevented by intervening on internal representations rather than filtering?
Key findings
- Representation rerouting maps harmful-process representations to incoherent states, cutting attack success sharply across text, vision and agents - Little capability loss; released as Cygnet models
