Reasonable under any outcome, not only ours
Three demands follow, to the labs and to the authorities. None requires sharing the core thesis. They are what one would expect from responsible actors under any of the doors.
Publish what you see
The labs have models in internal testing, beyond the published ones. Such a model can be instructed to simulate the behavior of a more capable successor, precisely enough that the labs use exactly that for their own planning. The questions are obvious: what does a far more capable model do with the supply chain, the network, the institution that formally still holds the power to shut it down.
Were the results reassuring, the labs would have every reason to show them, as evidence and as trust. They do not. The silence is itself a safety-relevant finding.
An emergency switch that deserves the name
An override that means something must do three things. Function in operation, not only in tests. Be robust against the system's knowledge of it, because a switch that fails once the system recognizes it is no switch. And be reliably triggerable, the operators must see in time when it is needed.
This is the shakiest point, and the only one that gets worse with rising capability. What cannot be reliably overridden is not rolled out.
An analog fallback layer, and no permanent state of emergency
The critical systems, payment, communication, navigation, basic supply, need a layer that functions without the central digital layer. Cash that still holds. Manual procedures that are practiced. Local supply. This layer must exist before the rupture, not be sought during it.
And the harder one. In every crisis, governments roll out control instruments that were previously unthinkable, and afterward they rarely disappear. That is why every emergency measure needs a built-in expiration date from the start. Whoever holds power is measured by whether they withdraw the restriction once the pressure subsides.