Early work on monitorability evaluations

Introduction Future AI systems may be capable enough to carry out sabotage, either via malicious real-world actions or via intentionally tampering with evaluations. We want to track the sabotage…