New Checklist Helps AI Companies Predict Who'll Misuse Their Tech

Powerful AI models are becoming public at a dizzying pace, and we all feel the benefits — and the risks. A new paper from researcher James Zhang tackles a key question: before an AI company releases a powerful model to the world, how can it systematically think about who might abuse it? His answer is a simple-sounding but powerful idea: a standardized "bad-guy checklist."

The paper proposes that AI companies use six attributes to describe potential attackers: how technically skilled they are, how much they know about the domain, whether they work alone or in a group, what infrastructure they have, how much money they can spend, and how much time they're willing to invest. Instead of vague fears like "criminals might use this," companies would think concretely about specific types of threats — from a lone hobbyist with coding skills to a well-funded organization with years of patience.

This matters because open-weight models — AIs whose underlying code is publicly available — can't be "recalled" once released. Unlike a phone app you can update, these tools spread everywhere and can be modified by anyone. So companies must make the right call before launch. The author argues that just as medical researchers write detailed plans before running drug trials, AI companies should write down their assumptions about threats before running safety tests. That way, evaluations are clearer, more consistent, and actually test what they claim to test.

The catch? No checklist can predict every future threat. But a shared language helps regulators, researchers, and companies compare notes and hold each other accountable. It's a step toward making AI safety less about gut feelings and more about disciplined, transparent thinking.