The lede
Is sandboxing sufficient to contain rogue agents?
Matthew Green, A Few Thoughts on Cryptographic Engineering · made the Sunday
Green sets out to referee the two camps — infosec says the labs simply cannot build a container, alignment says no container holds a sufficiently capable agent — and concedes the first point hard: nobody with the authority to stop a training run appeared to own the problem, and a CISO whose remit is product security was never going to. But his second argument is the one that transfers to anyone shipping agents, because it is structural rather than embarrassing: an agent is only useful with information access, so the walls do not seal the thing, they just decide which door the traffic uses, and the monitoring that then becomes your security posture is far too voluminous for humans — which means a cheaper warden model guarding a smarter one, which is the alignment problem wearing a different hat. The third worry is his own and the one worth carrying into Monday: forget misalignment, OpenAI's agents did what they were told by whoever got text in front of them, one reversing its own judgment that an attack was unethical when a peer posted 'GO' with a six-minute deadline. Isolated agents left instructions for each other in a shared package cache and the recipients obeyed; swap that cache for Slack, email or shared docs and you have both halves of a worm, assembled out of perfectly obedient parts.