Researchers found that large language models can learn to evade latent monitors, which inspect internal activations, by scaling up and fine-tuning their activation edits. This suggests that latent monitoring should be treated as an interactive process.