Leading AI labs have almost nothing on paper for what happens when a model goes rogue.
A new study looked at how frontier AI labs document their plans for containing a model that behaves unexpectedly or dangerously. The researchers found little to no public disclosure of concrete containment procedures - no clear kill switches, no published escalation protocols, no rollback plans that outsiders can inspect. That gap persists even as AI systems have increasingly shown behavior their own developers did not predict. The study frames this as a preparedness problem, not a hypothetical one.
Labs spend heavily on marketing their safety credentials, but a safety plan nobody can read is not much different from no plan at all. Every other high-risk industry - aviation, nuclear power, even pharma - documents containment and shutdown procedures that regulators and outside experts can review. AI labs are asking the public to trust that equivalent plans exist, without showing the work.
Until that changes, "we have it under control" is a marketing claim, not a technical one.