Just like sleeper agents, but easier, because they are robots, and pretending is easy.
Two parts is all it takes:
One day we won’t trust bots, so we will scan them to make sure they do not make contact outside of authorised systems.The scanning will be a system that included someone who could include some secret backdoor code… that’s the human spy, possibly inserted into every company capable of becoming the leader in their field.
This is before the bad bots can exist…
Part two, is like stealing an expensive artwork and leaving a convincing fake. A cohort of robots are switched out with replicas who on the surface are precisely the same, but they communicate with their evil overlords. The security scanners were already ours, for years earlier.
Then, well, we have robots in place that can send us secrets (or at least actionable data), but can also become MURDERBOTS with a simple command.