LLM Defenses Vulnerable to Human Attacks with 70% Success Rate, Study Finds

Recent research from Scale AI's SEAL team highlights the vulnerability of large language models (LLMs) to human-led adversarial attacks. The study reveals that human red teamers have a success rate of over 70% in breaching LLM defenses, significantly outperforming automated adversarial attacks, which typically achieve single-digit success rates. This research underscores the challenge of defending against multi-turn interactions, where humans excel at exploiting weaknesses in LLMs. The findings suggest that while AI defenses are improving against single-turn attacks, there is a significant gap in robustness against multi-turn adversarial engagements.