Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller
We present Cancerous Backdoor Attack, that gradually corrupts an LLM's reasoning over a sequence of interactions by leveraging invisible tokens and a stage controller.
ProofPaper ↗
Key points
- Backdoor Attacks on Large Language Models (LLMs) typically follow a trigger-response pattern, where satisfying a trigger condition activates a single terminal behavior.
- We evaluate Cancerous Backdoor Attack across several reasoning datasets.
- We compare Cancerous Backdoor Attack with prior backdoor attack baselines to demonstrate our proposed attack outperforms the baselines.
- Our results indicate that reasoning corruption can develop gradually before it becomes visible, making answer accuracy alone insufficient for detecting backdoor attacks.
Sources (1)
- [1]Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage ControllerarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 07:05 PM
We present Cancerous Backdoor Attack, that gradually corrupts an LLM's reasoning over a sequence of interactions by leveraging invisible tokens and a stage controller.
Backdoor Attacks on Large Language Models (LLMs) typically follow a trigger-response pattern, where satisfying a trigger condition activates a single terminal behavior.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 8, 2026GeopoliticalQA: Benchmarking Retrieval and Generation over Intelligence Reports
- Oct 8, 2026Knowing Is Not Telling: Identity Ossification in LLM Representations of Trans-border Communities
- Oct 8, 2026STAPLE: Static Sparse Upcycling of Pretrained Checkpoints via Embedding Modules
- Oct 8, 2026Understanding Jev, the new model everyone is talking about
- Oct 8, 2026Launching an opt-in vulnerability-finding service for open-source software
- Oct 7, 2026browser-use/browser-use 0.13.11