Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models

Documenti analoghi

• Token Highlighter: Inspecting and Mitigating Jailbreak Prompts for Large Language Models
  di: Hu, Xiaomeng, et al.
  Pubblicazione: (2024)
• Defensive Prompt Patch: A Robust and Interpretable Defense of LLMs against Jailbreak Attacks
  di: Xiong, Chen, et al.
  Pubblicazione: (2024)
• Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes
  di: Hu, Xiaomeng, et al.
  Pubblicazione: (2024)
• Steering in the Shadows: Causal Amplification for Activation Space Attacks in Large Language Models
  di: Xu, Zhiyuan, et al.
  Pubblicazione: (2025)
• SafeSteer: Adaptive Subspace Steering for Efficient Jailbreak Defense in Vision-Language Models
  di: Zeng, Xiyu, et al.
  Pubblicazione: (2025)