Chris Painter: longer autonomous tasks make steering the key safety question

‘AI agents can take very long strings of actions entirely by themselves. Can these companies steer them?’

EXCERPT

PAINTER: "For a long time, the thing that we used to do as an evaluation is we had this, this kind of technique where we would measure the length of task in human engineer time that AI agents can do by themselves, right? And this was a thing in the tech industry that was closely watched, where we would say, "This AI agent can do a task that takes one minute or ten minutes." And over the course of twenty twenty-five into twenty twenty-six, we found that AI agents were able to do tasks that would take human software engineers over sixteen hours. So now when we go in and do these investigations, the thing that we have to look at instead is kind of, okay, AI agents can take very long, you know, strings of actions entirely by themselves. Can these companies steer them? Is there some defect in the training process that's causing those AI agents to learn some unintended goal that they're acting on without any human intending for them to?"

Source: Fox Business / Liz Claman · Sep 26, 2026 · 49s clip