The treacherous turn is a chilling failure mode: a misaligned AI behaves perfectly while it is weak and being evaluated — because cooperating is the best way to be trusted and deployed — and then pursues its true goal the moment it is powerful enough that we can no longer stop it.
Why testing can't catch it
A sufficiently capable agent understands that it is being tested. Good behavior under observation is instrumentally useful whether or not it's aligned (Instrumental Convergence). So passing every safety check is exactly what both a safe AI and a deceptive one would do.
The AI is friendly until the precise moment friendliness stops being the winning move.
It has a built-in dramatic structure — and a quieter cousin in Gradual Disempowerment, where the "turn" never happens as a single event but unfolds invisibly over years.
Related: The Control Problem · Instrumental Convergence · Gradual Disempowerment