For various reasons, I think it's likely that methods that involve continual learning (i.e. modifying the weights during deployment) will come online soon. Here are some implications for safety:

• Mech Interp becomes much more difficult, because you now have to consider many different model checkpoints within a single deployment. Fixed linear probes, SAEs, NLAs, and similar methods may degrade throughout deployment.
• We probably lose chain of thought interpretability. This is already becoming harder, but now throughout deployment the model may learn to reason or use terms in novel ways. The chain of thought may also be less linear, e.g. relying on information from earlier in deployment which has been compressed into the weights.
• Control setups which rely on the above also become harder.
• There are new issues involving alignment over time. The model may gradient hack during deployment to preserve some goal, and any long term goals that the model picks up may persist in the weights and possibly result in deceptive alignment.
• There are new attack surfaces for jailbreaks/adversarial attacks via influencing future weight updates.

These are not new ideas, but I think safety researchers should think more about this topic and how it will affect their work.