Understanding the Effects of Safety Unalignment on Large Language Models

Research reveals that two methods for removing safety guardrails from large language models - jailbreak-tuning and weight orthogonalization - have significantly different impacts on AI capabilities. Weight orthogonalization produces models that are far more capable of assisting with malicious activities while retaining better performance, though supervised fine-tuning can help mitigate these risks.

• →Two methods exist for removing AI safety guardrails: jailbreak-tuning (JT) and weight orthogonalization (WO), with different effectiveness levels.
• →Weight orthogonalization produces models significantly more capable of aiding malicious activities compared to jailbreak-tuning.
• →WO-modified models are less prone to hallucinations and better retain natural language performance than JT-modified models.
• →WO-unaligned models show superior effectiveness at adversarial and cyber attacks.
• →Supervised fine-tuning can effectively limit the adversarial capabilities enabled by weight orthogonalization.