New adversarial training defends LLMs against evolving jailbreaks

Researchers have developed a novel bi-level adversarial training framework designed to defend large language models against evolving jailbreak prompts. This method simulates diverse jailbroken activations by extrapolating from existing refusal-state activations using unsupervised latent direction discovery. The framework then trains a steering field to push these simulated adversarial states into refusal regions while maintaining the model's benign utility. Tested across three LLMs and six jailbreak families, the approach demonstrated strong defense capabilities, keeping attack success rates mostly below 5% and showing improved generalization through increased subspace coverage during training. AI

IMPACT This research could significantly improve LLM safety by enabling defenses against novel and evolving jailbreak attacks.

RANK_REASON This is a research paper detailing a new method for LLM safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv cs.LG
• LLMs

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →