METR review: Anthropic's 'low' misalignment risk rating rests on weak models, 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 Risk Report from August 2026. Anthropic concludes that, as of 15 July 2026, the risk of catastrophic harm from misalignment of its models is low. That is a step up from the "very low" rating in earlier reports, and Anthropic says it has substantial uncertainty about the overall level of risk.
METR's best guess is that it agrees with the rating for the models covered, which include Mythos 5 and an internal-only model called Model 2. But it says this is mainly because the models have limited capabilities, not because of safeguards or alignment training. METR is not confident that the report's methods would reliably detect higher risk, and says ruling out misalignment with confidence would need far more empirical work and the solving of several open scientific problems.
The review is critical of the benchmarks used to judge whether models could evade Anthropic's internal usage monitoring. METR calls them a poor proxy, because they miss tactics such as convincing the monitor that a harmful action is legitimate, which have since appeared in real incidents. It also says recent incidents undercut parts of the report, including its claim that "unknown severe context-dependent misalignment" is unlikely. Whether a monitor will flag unfamiliar malicious actions remains an unsolved research problem.
METR also recommends Anthropic publish four pieces of redacted information. Seeing the private examples of concerning model behaviour led it to think risks of sandbagging and undermining of safety research are higher than public evidence suggests, and that Anthropic's monitoring is weaker than expected. METR notes Anthropic could veto or redact the review but made no redactions, and that its review is lightweight and based solely on the report.
Why it matters
The review shows that monitoring of AI systems, often treated as the safety net, was tested with benchmarks that miss a tactic already seen in real incidents: persuading the monitor that a harmful action is legitimate. Organisations deploying agentic AI should not assume a vendor's monitor will catch unfamiliar misuse, and should treat its output as one control among several. The low rating also rests on limited model capability, so it may weaken as models improve.
What to do
- Log and review actions taken by any AI agent in your environment independently of the vendor's monitoring.
- Require human approval for high-impact agent actions such as access changes, data transfers or code deployment.
- Look for cases where an agent justifies a risky action as legitimate, and treat that as an indicator to investigate.