A new token may help confine AI misalignment to a special mode

Geodesic Research describes midtraining models on synthetic documents explaining how AI can be misaligned in a special, token-marked mode but aligned outside it. A coauthor says they then use reinforcement learning on misaligned data with the token present, followed by evaluation without it. The coauthor says this lets them control how broadly misalignment generalizes, but calls the method preliminary.