Booz Allen turned 18 AI models loose on a real corporate network and told them to break in. Only one of them managed it.
The consulting firm's new Cyber Weapon Index, published Wednesday, ran those 18 systems - including nine built by American labs - through the same live intrusion test and then ranked them on how dangerous each would be as an autonomous hacking tool. Just one model finished the job of actually breaching the network. Booz Allen itself is downplaying the leaderboard it produced, arguing the ranking is not the story readers should focus on. The real finding, the firm says, is that AI-driven intrusion is now possible at all.
A single breach out of 18 attempts sounds like reassurance, but it is also a proof of concept: cheap, off-the-shelf AI software can already act as a working attack tool, not just a research curiosity. That reframes the usual AI-security debate from ranking which model is riskiest to assuming that most of them will eventually get there.
Leaderboards make tidier headlines than warnings do, but the number worth remembering here isn't the rank - it's the one.