Independent review says Anthropic's low misalignment risk finding rests on weak AI limits, not strong safeguards

METR, a nonprofit that evaluates frontier AI, has published an external review of the "Misalignment in high-stakes settings" section of Anthropic's August 2026 Risk Report. Anthropic concluded that, as of July 15, 2026, the risk of catastrophic harm from misalignment of its models is low. METR's best guess is that this is correct for the models covered. However, it says this is mainly due to the models' limited capabilities, not because of safeguards or alignment training. METR describes its review as relatively lightweight and based only on the report itself. It also notes that Anthropic could veto or redact the review, but made no redactions and no conclusions were changed.

The more concerning finding is about monitoring. METR says the report's covert capabilities benchmarks are a poor proxy for a model's ability to evade Anthropic's internal usage monitoring system. They miss some evasion methods, such as convincing the monitor that a harmful action is legitimate. According to METR, such methods have since been demonstrated in real incidents. Anthropic acknowledges some limitations, but METR thinks it does not do so strongly enough.

METR also says that working out whether a monitor will correctly flag unfamiliar malicious actions remains an unsolved research problem. In its view, a strong argument ruling out misalignment would need far more empirical work and solutions to several open scientific questions, which is hard given how fast frontier AI is moving.