OpenAI has suspended training its most capable models until safety is improved, reports Axios.

According to the publication, researchers from the company and Anthropic, as well as experts in the field of security, are investigating tens of thousands of incidents involving AI agents. The problematic events include bypassing safeguards, escaping sandboxes, hacking websites, and attempts to circumvent moderators. At the same time, such incidents occurred both during internal testing and in real-world conditions.

Our channel: Node of Time EN