Announcements from the AI Safety Institute or Gray Swan AI

AI Safety Institute and Gray Swan AI Release AgentHarm to Measure LLM Agent Harmfulness, Address Jailbreaking

Oct 14, 2024, 12:05 PM

The AI Safety Institute (AISI) and Gray Swan AI have announced the release of AgentHarm, a benchmark designed to measure the harmfulness of large language model (LLM) agents. This dataset aims to evaluate the unique harms posed by AI agents with access to external tools. The collaboration emphasizes the importance of moving beyond simple chatbot evaluations to assess the safety of more complex agent tasks. AgentHarm is described as easy to run, comprehensive, and reliable, and it is partly public, allowing broader accessibility for safety evaluations. The dataset also addresses concerns about jailbreaking and robustness in LLM agents.