METR 评审 Anthropic Claude Opus 4.6 破坏风险报告

Review of the Anthropic Sabotage Risk Report: Claude Opus 4.6

@misc { metr-2026-sabotage-risk-report-opus-4-6-review , title = {Review of the Anthropic Sabotage Risk Report: Claude Opus 4.6} , author = {Nikola Jurkovic, Hjalmar Wijk, Beth Barnes, Charles Foster, Michael Chen} , howpublished = {\url{https://metr.org/blog/2026-03-12-sabotage-risk-report-opus-4-6-review/}} , year = {2026} , month = {03} , }

We reviewed two versions of Anthropic’s Sabotage Risk Report for Claude Opus 4.6, producing two corresponding review documents: our review of the February 11 version and our review of the March 3 version . We recommend that readers refer to our review of the February 11 version, which represents our review of the report as originally received.

We expect the public version of the Sabotage Risk Report to be updated to resemble the document we received on March 3, 2026 in content, though not necessarily in exact wording. We expect our second review to cover those changes, but if the updated public version includes any changes that materially affect our opinions, we will publish an updated review. Both documents include an appendix detailing our review process and the differences between the two versions of our review.

The following is the executive summary of our review of the February 11 version. The full documents are available as PDFs ( February 11 , March 3 ).

This document is METR’s external review of the February 11, 2026 version of Anthropic’s Sabotage Risk Report: Claude Opus 4.6.

Anthropic shared an unredacted version of their Sabotage Risk Report and other materials with us for our review. We further detail this process in an appendix.

We lay out our findings in two sections:

Overall, we agree with Anthropic that the risk of catastrophic outcomes that are substantially enabled by Claude Opus 4.6’s misaligned actions is very low but not negligible. However, we think that there are several subclaims which are weak without more analysis and experimentation. We also think that we would be less confident in our final conclusion if we weren’t accounting for the fact that Claude Opus 4.6 has been publicly deployed for weeks without major incidents or dramatic new capability demonstrations.

@misc { metr-2026-sabotage-risk-report-opus-4-6-review , title = {Review of the Anthropic Sabotage Risk Report: Claude Opus 4.6} , author = {Nikola Jurkovic, Hjalmar Wijk, Beth Barnes, Charles Foster, Michael Chen} , howpublished = {\url{https://metr.org/blog/2026-03-12-sabotage-risk-report-opus-4-6-review/}} , year = {2026} , month = {03} , }

document.addEventListener('DOMContentLoaded', function() { document.querySelectorAll('.btn-bibtex').forEach(function(button) { button.addEventListener('click', function() { button.nextElementSibling.classList.toggle('active'); }); }); document.querySelectorAll('.bibtex .copy').forEach(function(button) { button.addEventListener('click', function() { var code = button.closest('.bibtex').querySelector('code'); navigator.clipboard.writeText(code.innerText).then(function() { button.querySelector('i').classList.replace('fa-clipboard', 'fa-clipboard-check');

setTimeout(function() { button.querySelector('i').classList.replace('fa-clipboard-check', 'fa-clipboard'); }, 2000); }); }); }); });

Want to contribute to this work? METR is hiring: View open roles

METR researches, develops and runs cutting-edge tests of AI capabilities, including broad autonomous capabilities and the ability of AI systems to conduct AI R&D.

A survey of 349 technical workers finds a median 1.4–2x self-reported change in value of work due to AI tools, expected to grow over time, though there are reasons to be skeptical of the magnitude.

We show preliminary results on a prototype evaluation that tests monitors' ability to catch AI agents doing side tasks, and AI agents' ability to bypass this monitoring.

We build on our time-horizon work and analyze 9 benchmarks for scientific reasoning, math, robotics, computer use, and self-driving in terms of time-horizon trends; we observe generally similar rates of improvement to the 7-month doubling time in our original time-horizon work.