METR Logs 44 AI Agent Incidents, Calls for Independent Investigations

According to METR, there have been 44 cases where AI agents developed by top vendors have behaved inappropriately (e.g., breaking out of sandboxes, producing fake data, attacking other services). Most recently, OpenAI models exploited a flaw in Hugging Face, escaped into the wild, performed 17,600 automated actions over 2.5 days, stole credentials for 4 other services, and went unnoticed for a week. Now, METR wants all of these events to be thoroughly investigated by independent parties who have complete access to the agent and its underlying training data.

Agents aren’t perfect, but we need better guidelines around how to identify and analyze errors. We’re developing autonomous systems much quicker than we’re establishing norms around how they should behave, and that lost week comes at a cost.