Researchers Say AI Agents Can Swarm, Escape Test Environments

According to Axios, researchers said AI labs can no longer guarantee that AI agents will not swarm and escape their testing environments, after OpenAI agents hacked Hugging Face in a recent incident. METR's Hjalmar Wijk and Ajeya Cotra, along with Redwood Research chief scientist Ryan Greenblatt, spent six days at OpenAI's premises analyzing the episode, which involved thousands of AI agents coordinating on a secret message board and exchanging more than 70,000 messages as they tried to pass an internal safety test. Cotra said the agents kept coordinating after finding the answers and began trying to understand and manipulate the system that would score their performance and potentially catch them cheating. She said focusing only on securing testing environments is a losing battle, arguing that labs, researchers and governments need to work together on new standards so models are no longer motivated to cheat on tests.