On May 19, METR, an AI safety assessment organization, released its first-ever Frontier Risk Report (covering the evaluation period from February 16 to March 16, 2026). The report reveals findings from METR’s in-depth entity-level risk assessments of four leading AI labs: Anthropic, Google, Meta, and OpenAI. METR was granted access to each lab’s most powerful internal models (including their raw reasoning chains), along with extensive non-public information regarding model capabilities, internal usage patterns, and development progress. The assessment framework relies on three core dimensions—means, motives, and opportunities—and specifically focuses on the risk of internal AI agents carrying out ‘rogue deployments’, meaning agents operating autonomously without human authorization. The report concludes that during the evaluation period, internal agents at all four labs ‘likely possessed the means, motives, and opportunities required to conduct small-scale rogue deployments’, though they lacked the robustness needed to remain undetected or uncorrected. METR also warns that given the rapid pace of AI advancement, the robustness of such rogue deployments ‘is expected to increase significantly in the coming months’, and plans to conduct similar assessments again by the end of 2026.
Four days after the report’s release, Elizabeth Barnes, founder and CEO of METR, posted a lengthy thread on X on May 23. Outside the scope of the report’s framework, she shared her personal views on the broader AI safety landscape: (1) ‘We are likely on track to develop AI systems capable of causing human extinction or permanent loss of autonomy, possibly within the next few years’ (Barnes later clarified that this refers to the prerequisite condition of AI possessing such ‘capability’, rather than a direct assertion that disaster will occur); (2) ‘The situation is chaotic and rushed; we haven’t even addressed basic issues’; (3) ‘Independent organizations like METR are severely under-resourced relative to the speed of AI development’; (4) ‘Any rational civilization would have slowed down long ago’. Ryan Greenblatt, a researcher at Anthropic, and Miles Brundage, an AI safety researcher, publicly expressed agreement with her stance. Jackson Kernion, an Anthropic employee, questioned the specific mechanisms behind AI ‘takeover’, to which Oliver Habryka, a researcher at MIRI, pointed out that Anthropic was founded precisely to prevent such risks from materializing. Within 24 hours, this thread garnered over 680,000 views, sparking widespread discussion across the AI safety community.