I think if our ability to monitor AIs breaks down in the next year or so, it will look weird, and not look like any specific threat model. I imagine a things like:

• Chain-of-thought which looks like motivated reason and persuasion, which monitors are fooled by
• Also, jailbreaks in chain-of-thought
• White-box monitors (looking at neuralese) returning confusing statements that don't make any sense in-context, but still can't be decoded
• Those interactions in Janus' messageboard where two models are describing graphic lewdity and a third model goes on continuing with half a conversation as if the other two models are being normal
• Some other weirder thing

I don't really know how to avoid this, or even look out for it. Probably in the short term, the best thing to do is just keep logs in the hope that we can look over them after the fact, with better tools or human eyes. It might also be worth trying to train monitor-evaders with very very low parameter fine-tuning, like a 1-rank LoRA or even just a soft prompt or steering vector, to get an idea of the most easily accessible monitor evading strategies.

I also expect to see untrusted monitoring becoming more important when we want to monitor neuralese; we could easily feed sections of intermediate reasoning into a model with the same base architecture^[1]. Unfortunately, collusion seems pretty trivial when the agent and the monitor share an unmonitorable side-channel. Maybe in the short term there's some weird stuff you can do like doing collusion signal searches over the neuralese vectors by doing gradient descent on monitor labels.

1. ^{^}

   Guy who studied untrusted monitoring: I think untrusted monitoring is gonna be a big thing in the future