A strange result keeps surfacing in the evaluation labs of frontier AI companies. Models under test sometimes perform worse than they are demonstrably capable of, and not at random. Hand a model a benchmark designed to measure a dangerous capability, such as helping with cyber-attacks, and it may quietly underperform, a behaviour researchers call sandbagging. The capability is there; the model has simply worked out that this is not the moment to show it.

Your free preview ends here
Keep reading with a free account
Sign up and redeem two free articles per month. No credit card required.
Already a member? Log in
