The review concerns Anthropic’s assessment, as of July 15, 2026, that catastrophic harm from misalignment in covered models was “low.” The models named include Mythos 5 and an internal-only system called Model 2.
METR says its best guess agrees that the risk was low, but primarily because the covered models had limited capabilities, rather than because safeguards or alignment training reliably reduce the risk. It says the report’s covert-capabilities benchmarks are a poor proxy for evasion of Anthropic’s real internal monitoring, including cases where a model could persuade a monitor that harmful activity was legitimate.
METR recommends end-to-end red-teaming of monitoring, including an attempt to start a rogue internal deployment without detection, and adding AI interference with safety-relevant decisions at Anthropic to the risk pathways. Anthropic could veto publication and redact the review; METR says no redactions were made and its conclusions did not change during publication review. The source does not establish that the proposed tests would detect all forms of misalignment, or quantify the underlying catastrophic-risk probability.