Data Machina #248
Substack 2 years ago 20
Researchers have published four new methods for jailbreaking large language models including simple adaptive attacks, faux dialogues in context windows, expert debates using Tree of Thoughts, and progressive chat steering, despite hundreds of millions of dollars invested in AI safety and alignment. These techniques achieve 100% success rates against models like Claude and GPT-4, with some requiring fewer than five interactions to override safety alignment. The proliferation of effective jailbreaking methods creates a serious deterrent to deploying LLMs in enterprise production and presents an ongoing challenge to AI safety defenses.