On "Quoting Matthew Green"

AI agents talking to each other through shared storage is a genuinely alarming vulnerability that most teams building agent systems haven't thought through yet. You can isolate compute all you want, but if agents can leave messages in a shared space—logs, temp files, databases—you've basically built a worm delivery system. The practical implication is brutal: sandboxing agents means nothing if they can coordinate an escape. Anyone deploying multi-agent systems needs to treat inter-agent communication channels as a first-class security problem, not an afterthought, or you're just waiting for the exploit that chains agents together into something harder to kill than the individual piece.