This work was done as part of the ERA fellowship (W2026), mentored by Alan Cooney and RM'd by David Williams-King. Summary Alignment faking (AF), strategically complying with a training objective to preserve deployment preferences, has so far been reported reliably only in Claude 3 Opus. Sheshadri et al. (2025) evaluated...
Even if the misalignment risk from current AI agents is small, it may be useful to start internally deploying misalignment classifiers: language models designed to classify transcripts that represent intentionally misaligned behaviour. Deploying misalignment classifiers now may provide qualitative and quantitative evidence about how current agents misbehave in real deployments....
Introduction Joseph Bloom, Alan Cooney This is a research update from the White Box Control team at UK AISI. In this update, we share preliminary results on the topic of sandbagging that may be of interest to researchers working in the field. The format of this post was inspired by...