CIAware-Bench authors discuss new results on control intervention awareness evaluation

Alexander Panfilov discusses CIAware-Bench's control intervention awareness evaluation, reporting that recent frontier models perform better in the tested settings. He explicitly states that the current tests only examine whether models can identify interventions when explicitly asked. The results cannot directly imply that models would spontaneously notice interventions, the occurrence rate in real deployments, or that all control methods universally fail.