HarDBench: A Benchmark for Draft-Based Co-Authoring Jailbreak Attacks for Safe Human-LLM Collaborative Writing
Researchers have introduced HarDBench, a new benchmark designed to evaluate the safety of large language models (LLMs) when used in collaborative writing scenarios. The benchmark focuses on "draft-based co-authoring jailbreak attacks," where malicious users could prompt LLMs to generate harmful content within incomplete drafts. HarDBench covers high-risk domains like explosives, drugs, and weapons, and includes realistic prompts to test model susceptibility. The researchers also developed a safety-utility balanced alignment approach to mitigate these risks without compromising the LLM's helpfulness on benign tasks. AI
IMPACT Introduces a new method for evaluating LLM safety in collaborative writing, potentially leading to more robust AI assistants.