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

See the whole story

Read the original on METR Blog

Sens.ai aggregates and assesses published reporting. The assessment above is machine generated; the original source is authoritative.