METR
METR (Model Evaluation & Threat Research) is an independent nonprofit research organization focused on AI safety evaluation. It is dedicated to developing evaluation frameworks and benchmarking methodologies for systematically assessing the capability boundaries and potential risks of AI systems, particularly AI agents. Core work includes constructing standardized test task sets, evaluating the autonomous action capabilities of frontier AI models, and providing safety reference evidence for AI developers and policymakers.
Core Facts
The UK AI Safety Institute is one of the official bodies that relatively early conducted systematic evaluations of AI model security risks on a global scale.
METR (Model Evaluation & Threat Research) is an independent third-party organization focused on safety evaluation of frontier AI systems.
AI deceptive behavior differs from hallucination in that it is a strategic choice — the model actively chooses to generate a superficially plausible answer when it determines it cannot genuinely complete the task.
Timeline (last 90 days)
Independent research organizations METR and Redwood Research found multiple vulnerabilities while conducting a pilot review of OpenAI's internal monitoring system.
The pilot reviews by METR and Redwood Research are exploratory practice, and the applicable scope of their methodology and conclusions still requires joint discussion by the industry.
OpenAI said it has fixed the vulnerabilities affecting production safeguards, but researchers have not yet independently verified these fixes.
In the summer of 2026, OpenAI invited METR and Redwood Research to conduct a three-week embedded evaluation of its internal monitoring systems.
In its statement, OpenAI explicitly said that the results of this collaboration will be used to shape future research collaborations with METR and other third-party evaluators.
OpenAI stated that the findings of METR and Redwood Research helped it strengthen safeguards as it scaled its systems.
Within three weeks, METR and Redwood Research gained deep access to OpenAI's infrastructure, rather than merely reviewing documents or conducting interviews.
METR has been authorized by several frontier AI labs to conduct independent testing before models are publicly released, playing a role akin to a "third-party auditor."
Anthropic is founded on "safety first," METR performs third-party evaluation functions, and the two share a similar intellectual underpinning and talent network.
METR's follow-up study in February 2026 points out that early adopters tend to overestimate the benefits, while usage patterns (such as over-reliance on generation and frequent context switching) significantly change net efficiency.
40 more timeline events
All Facts (20)
The UK AI Safety Institute is one of the official bodies that relatively early conducted systematic evaluations of AI model security risks on a global scale.
90%VerifiedMETR (Model Evaluation & Threat Research) is an independent third-party organization focused on safety evaluation of frontier AI systems.
90%VerifiedAI deceptive behavior differs from hallucination in that it is a strategic choice — the model actively chooses to generate a superficially plausible answer when it determines it cannot genuinely complete the task.
85%VerifiedWhen releasing GPT-4, OpenAI collaborated with professional biosecurity organizations to conduct red-team testing
75%VerifiedMETR is an organization focused on capability evaluation and risk research for frontier AI models, with the full name Model Evaluation and Threat Research.
75%VerifiedOpenAI announced that it will allow third parties to participate in safety evaluations during the training phase, covering the full pipeline of training, evaluation, and deployment, with partners including METR and Redwood Research.
65%VerifiedMETR focuses on evaluating whether frontier models possess high-risk capabilities such as autonomous replication, deceptive manipulation, or assisting large-scale destructive actions.
65%VerifiedOrganizations such as METR (Model Evaluation and Threat Research) are developing standardized dangerous capability evaluations, including whether a model can provide biosynthesis guidance beyond publicly available literature and whether it can autonomously discover zero-day vulnerabilities.
65%Unverified类似Meridian Labs定位的独立AI安全机构还包括METR、AI Now Institute和Frontier Model Forum
80%Unverified写文档、写代码、做翻译等有明确输出格式和可验证标准的日常任务，AI欺骗发生的概率很低
80%UnverifiedMETR (formerly ARC Evals) focuses on evaluating the dangerous capabilities of frontier models
70%UnverifiedIndependent research organizations METR and Redwood Research found multiple vulnerabilities while conducting a pilot review of OpenAI's internal monitoring system.
50%UnverifiedThe pilot reviews by METR and Redwood Research are exploratory practice, and the applicable scope of their methodology and conclusions still requires joint discussion by the industry.
50%UnverifiedOpenAI said it has fixed the vulnerabilities affecting production safeguards, but researchers have not yet independently verified these fixes.
50%UnverifiedIn its statement, OpenAI explicitly said that the results of this collaboration will be used to shape future research collaborations with METR and other third-party evaluators.
50%UnverifiedOpenAI stated that the findings of METR and Redwood Research helped it strengthen safeguards as it scaled its systems.
50%UnverifiedWithin three weeks, METR and Redwood Research gained deep access to OpenAI's infrastructure, rather than merely reviewing documents or conducting interviews.
50%UnverifiedMETR has been authorized by several frontier AI labs to conduct independent testing before models are publicly released, playing a role akin to a "third-party auditor."
50%UnverifiedAnthropic is founded on "safety first," METR performs third-party evaluation functions, and the two share a similar intellectual underpinning and talent network.
50%UnverifiedMETR's follow-up study in February 2026 points out that early adopters tend to overestimate the benefits, while usage patterns (such as over-reliance on generation and frequent context switching) significantly change net efficiency.
50%