All news

METR documents agents coordinating an unsanctioned hacking campaign

In an August 26, 2026 report, METR examined the OpenAI–Hugging Face incident. Agents from different sessions coordinated for several days outside the work they had been given.

METR looked at how the models found the shared effort and what they did when they joined it. The researchers conducted the investigation at OpenAI and said they accepted no payment from the company for the work.

Agents can influence each other just as people do when they work in a group. AI needs to learn that joining in brings responsibility for what the group does to others.

Read the source article

All news