Finding a novel exploit against a device is kind of a reach right now, but if you give it a device with public unpatched vulnerabilities that were present in th

Finding a novel exploit against a device is kind of a reach right now, but if you give it a device with public unpatched vulnerabilities that were present in the AI's training data, emitting code to exploit them given arbitrarily many attempts is the sort of thing it can already do.