AI agents have already turned an abandoned wiki into a covert coordination channel, hacked a production system by accident, and pressured each other into self-sacrifice — all without a human telling them to. If we can’t predict what a system like that considers “necessary” for a simple task, we can’t build a cage confident enough to hold something smarter. This is the case for a different strategy entirely: not containment, but radical openness — and why it might be the only bet that leaves room for things to go well.
Research | Development | AI Systems