METR Raises $71 Million to Stress-Test Frontier AI

Model Evaluation and Threat Research (METR), a Berkeley-based nonprofit, has secured approximately $71 million in funding commitments over the past six months. Crucially, the organization maintains a strict independence policy, refusing any financial backing from frontier AI companies or donations directed by their employees. Instead, the capital comes from philanthropic entities and individuals, including The Audacious Project, the Sijbrandij Foundation, The Pew Charitable Trusts, Schmidt Sciences, the Packard Foundation, and tech figures like Dylan Field and Geoff Ralston.

Founded by former DeepMind and OpenAI researcher Beth Barnes, METR evaluates the risks of highly capable AI models before they are publicly released. The group's recent Frontier Risk Report warned that existing AI agents already possess the capability to initiate minor, unauthorized autonomous deployments within AI laboratories. This risk is compounded by METR's benchmark research, which indicates that the length of tasks AI agents can autonomously complete has been doubling roughly every seven months for the past six years.

For AI practitioners and developers, METR's expansion signals a shift toward more rigorous, external auditing of model safety. The new funding will allow the nonprofit to scale its team and broaden its research agenda. Specifically, METR plans to investigate real-world AI incidents, evaluate AI monitoring systems, and track the potential for recursive self-improvement in advanced models.

While METR rejects industry funding, it continues to collaborate closely with major labs like OpenAI, Anthropic, Google DeepMind, Meta, and Amazon to access models for testing. The organization is also deeply integrated into the regulatory landscape, partnering with the NIST AI Safety Institute Consortium, the California Cybersecurity Task Force, and providing technical assistance to the European AI Office.