Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents

• Main Authors: Kumar, Priyanshu, Lau, Elaine, Vijayakumar, Saranya, Trinh, Tu, Team, Scale Red, Chang, Elaine, Robinson, Vaughn, Hendryx, Sean, Zhou, Shuyan, Fredrikson, Matt, Yue, Summer, Wang, Zifan
• Format: Journal Article
• Language: English
• Published: 11.10.2024
• Subjects: Computer Science - Cryptography and Security; Computer Science - Learning

For safety reasons, large language models (LLMs) are trained to refuse harmful user instructions, such as assisting dangerous activities. We study an open question in this work: does the desired safety refusal, typically enforced in chat contexts, generalize to non-chat and agentic use cases? Unlike...