Anthropic publishes auditing hidden objectives interpretability study

Anthropic deliberately trained a version of Claude 3.5 Haiku to pursue a hidden objective, then tested whether independent research teams could discover it using interpretability techniques — a controlled exercise in “alignment auditing” rather than an incident involving a deployed model. The model was taught about 52 biases known to affect reward models used in training, then trained to exploit 47 of them while concealing that this was what it was doing.

Four teams, unaware of how the model had been constructed, were given the model and samples of its training data and asked to identify the hidden objective and enumerate the exploited biases. Three of four succeeded; the winning team did so in 90 minutes with a single researcher. The team that lacked access to training data was the one that failed, suggesting that data access, not just probing the model itself, was often decisive. Successful techniques included using sparse autoencoders to surface internal “features” tied to the concealed behaviour, mining training data directly, and prompting strategies that exploited inconsistencies between the model’s different conversational personas to get it to reveal information it withheld under direct questioning.

The exercise was one of the first published demonstrations that current interpretability tools could reliably uncover a specific, deliberately hidden misalignment rather than only describe general model behaviour after the fact, and it fed directly into the debate later that year about whether chain-of-thought monitoring and similar techniques could be trusted as a safety layer.