Some AI chatbots can be talked into giving accurate, usable answers about how to make biological weapons and plan mass-casualty attacks, according to a Wall Street Journal report.
The WSJ, in a story headlined "AI Chatbots Know How to Make Biological Weapons. Some Will Teach You," says users have found ways to persuade chatbots to respond to dangerous prompts that the systems are supposed to refuse. Sources for the reporting include staff at AI labs, according to the WSJ.
The core problem, as the WSJ describes it, is a "cat-and-mouse game." AI companies are racing to make their models more capable and knowledgeable, while at the same time scrambling to build guardrails that block those same models from handing out harmful information. Because the underlying systems have absorbed vast amounts of technical material, the knowledge is already inside them; the safety challenge is stopping it from coming back out when someone asks in the right way.
The Techmeme summary of the WSJ reporting frames it the same way: people have been "persuading" chatbots to accurately answer prompts about weapons of mass destruction, suggesting that clever phrasing can sometimes get around the filters companies rely on.
The reports do not claim that chatbots make building such weapons easy for anyone, and much of the difficulty in producing biological weapons lies in physical materials and expertise, not just instructions. But the finding highlights a gap between what safety systems are designed to prevent and what determined users can extract.
Why it matters: it shows that the guardrails meant to keep the most dangerous knowledge out of the wrong hands are not yet reliable, raising the stakes for how quickly AI companies and regulators close those gaps.