Anthropic's August 2026 risk report rates full automation risk low
Anthropic just released its August 2026 Risk Report, taking a close look at how risky its artificial intelligence (AI) systems really are, especially when it comes to automating research and handling sensitive stuff like chemical or biological weapons.
Right now, they're calling the risk of fully automated research "low," but admit they're less sure about that label as their AIs keep getting smarter.
Anthropic CoBench shows models below 85%
Their Claude Mythos 5 and Model 2, are helping speed up research but aren't ready to replace humans yet.
On its internal CoBench evaluation, Model 2 scored 62.8% and Claude Mythos 5 got 50.3%, both under the full automation mark of 85%.
Since early last year, AI has made things faster at Anthropic (though not quite doubled), and big fields like biotechnology and robotics haven't seen major AI takeover yet.
So for now, Anthropic says careful monitoring is still a must.