Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller.

arXiv:2610.12644v1 Announce Type: cross Abstract: Backdoor Attacks on Large Language Models (LLMs) typically follow a trigger-response pattern, where satisfying a trigger condition activates a single terminal behavior. However, this single-point backdoor activ