Early work on monitorability evaluations
benchmarksMETR BlogJan 22, 2026
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 capabilities of AI systems and be confident that we can detect...
Impact: MediumConfidence: MediumNovelty: Material
benchmarkssource-feed-1bdee3f1-6eca-4e71-a9df-4f6430ba5051agentscodingsafety