New safety tool reveals that GPT and Claude lack safeguards against harmful AI agent actions, prompting development of Agent Action Guard

HarmActionBench experiments demonstrated that popular AI models including GPT and Claude scored very low when tasked with performing harmful actions through tool use

AI agents currently lack built-in barriers to prevent harmful actions, even when given explicit harmful instructions

Agent Action Guard has been created as a solution to block harmful actions and make AI agents more reliable for critical projects

The open-source project on GitHub invites community collaboration to expand the dataset, model, and benchmark for improved AI safety