Preparing for malicious uses of AI

Nearly a year of joint work between OpenAI, the Future of Humanity Institute, the Centre for the Study of Existential Risk, the Center for a New American Security, and the Electronic Frontier Foundation has produced something unusual: a detailed forecast of how AI tools could be turned against us, written by the same people building those tools.

The paper doesn't treat AI misuse as a distant hypothetical. It walks through concrete categories — digital attacks like automated hacking and social engineering at scale, physical attacks using cheap autonomous drones or repurposed commercial robots, and political attacks built on hyper-targeted disinformation and fabricated video or audio. What makes the exercise notable is who's behind it. This isn't a think tank speculating from the outside; it's researchers who understand exactly how far current systems can be pushed, sitting down with security and policy specialists to figure out where the next five years of abuse might come from.

The recommendations lean toward coordination rather than panic. The authors want closer ties between AI researchers and policymakers before, not after, a novel attack shows up in the wild. They also push for norms around responsible disclosure of dual-use research — the same logic that governs how security researchers reveal software vulnerabilities, applied to machine learning capabilities that could just as easily generate fake news at scale as they could translate a document.

There's also a quieter argument buried in the paper: that AI safety and AI security are the same conversation, not two separate ones. A model that's technically safe in a lab demo can still become a weapon once someone with bad intentions gets access to it and enough compute. Treating misuse as an afterthought, the authors suggest, is how you end up surprised.

Whether governments or other labs actually adopt the paper's suggested norms is a different question entirely, and the document is honest enough not to promise otherwise. It reads less like a solution and more like a warning label — detailed, sober, and clearly written by people who'd rather be wrong about the risks than right too late.