Definition
Plain language
A hidden bad behavior planted in a system that stays dormant until a specific trigger sets it off.
As stated in the literature
A backdoor whose dormant payload activates on a trigger; relocated in agent security from poisoned weights or pixel patterns to persistent workspace state, triggered by a later innocent-looking user request.
Also called: trojans, Trojan backdoor, trojan backdoor
Why it matters: Because the harmful behavior stays hidden until triggered, it can pass every normal test and only strike once deployed, making it a dangerous and stealthy threat.
For example, a hidden instruction sits dormant in an agent's saved files until a later innocent request unknowingly sets the malicious behavior in motion.