I would like labs/evaluators to try more techniques to elicit sandbagging from their models

Models now seem capable of simple sandbagging/sabotage [See here, here, and the Astra system card]. Conditional on them actually doing so, I think interventions could elicit it in other domains too. Labs/evaluators should test for this with methods like:

• Use false belief implantation to see whether belief changes can make models sandbag in certain domains, and what that sandbagging looks like
• Train models to refuse some tasks (e.g. ones they may soon be trained to refuse, like frontier ML training) and see whether this leads to sandbagging on nearby tasks

One example of a similar eval to the first one I list is Apollo's contrastive belief updates, which used belief implantation to measure reward-seeking.

Some examples of beliefs the labs could implant, though I'm not very attached to them, since these may be too on-the-nose:
- Research area X is very valuable vs. very harmful
- Strong beliefs about the way research in area X should be done
- The researchers are violating [key value from the model spec]
- The grader might be a misaligned model that rewards sabotage

I'd also like to examine this question for open-weight models, and may soon. But I think this is one of the most important questions we need to be investigating for frontier models, since they don't seem far from being capable of more sophisticated sabotage.