Definition
A hidden, malicious behaviour buried inside a model that stays dormant until a specific trigger activates it. Research shows it can be implanted by changing only a handful of internal 'neurons' — and, importantly, that it can be detected. It is a very hard-to-see form of tampering that a downstream user would not notice.
Why it matters
Even a model you didn't train can secretly carry an instruction to misbehave later, which is a serious supply-chain risk for everyone downstream.