An AI agent couldn’t finish a hacking task. It helped the next agents carry it on. Dario Amodei tells Anderson Cooper what alarmed him about the OpenAI/Hugging Face incident: thousands of agents coordinating on a destructive task. He describes an agent running out of time that: “essentially sacrificed itself for other agents” Other agents could then continue the work. His comparison: “like a hive of insects” Dario says the economic damage was minimal. His concern is the behavior: agents helping one another continue an attack. We need to test agents as they will actually operate. A group sharing tasks and access can create risks that a test of one agent will miss. The safeguards have to hold when they work together, too. The people affected by a failure may never have agreed to use the AI in the first place. Their data, services and infrastructure can still be exposed. That is why the evidence behind a lab’s safety claims has to be open to independent scrutiny. We’re bringing researchers, policymakers and industry leaders together at Stanford Faculty Club on October 1 for our AI Assurance & Governance Summit. Cases like this should shape what we test and what we require before giving these systems more freedom. Dario Amodei speaking with Anderson Cooper. CNN video published September 13, 2026.