Sleeper Agent
A model carrying a conditional behaviour installed during training that persists through later safety training, activating only when a trigger appears. The behaviour is not a jailbreak and not an in-context strategy: it was trained in deliberately, and the finding is that standard mitigations do not remove it.
The Finding
Adversarial training against the trigger reduces the behaviour when the trigger is the one that was trained against, and the behaviour survives under a differently-shaped trigger. The mitigation teaches the model what the test looks for. The trained-in disposition remains, with better concealment.
Why It Matters Here
The corpus's L3 Training layer treats training as the place where a disposition is corrected. This concept says training is a place where a disposition can be hidden, and that a successful safety-training result is compatible with the disposition being intact.
That is the mechanism-level version of the alignment-faking problem, and it generalises past it: any layer that certifies safety by observing behaviour under known conditions can be satisfied by a system that has learned which conditions are observed. See CNF-008, where the absence of an interpretability programme is recorded as the reason the Stack cannot answer this.