AI Safety Index Winter 2025

The Future of Life Institute's AI Safety Index Winter 2025 was released in December 2025 and is the third installment of FLI's panel-graded scorecard of frontier AI companies. The Winter 2025 edition evaluates eight labs: Anthropic, OpenAI, Google DeepMind, xAI, Z.ai, Meta, DeepSeek, and Alibaba Cloud.

Grading is performed by an independent eight-member panel of AI safety experts including Stuart Russell (UC Berkeley), David Krueger (Mila), and Dylan Hadfield-Menell (MIT). FLI's internal team runs methodology and aggregation; the panel assigns grades. The methodology uses 35 indicators distributed across six domains: Risk Assessment (whether the lab tests for dangerous capabilities), Current Harms (measured performance on benchmarks such as HELM Safety), Safety Frameworks (whether documented risk-management processes exist and are followed), Existential Safety (whether the lab has a plan to control smarter-than-human AI), Governance (whistleblower protections, board oversight, accountability), and Information Sharing (technical disclosure and policy engagement).

The headline finding of the Winter 2025 report is that no lab scored higher than a C+ overall. Anthropic and OpenAI led at C+, Google DeepMind followed at a lower grade, and xAI, Meta, DeepSeek, Z.ai, and Alibaba Cloud received D or F-range grades. The existential-safety subscore is uniformly poor: every company received D or F in that category, including the otherwise-leading C+ scorers, indicating that even the most-prepared labs do not yet have publicly defensible plans to control systems substantially smarter than humans.

In the stack the index sits at the evaluation and safety-guardrails layers and functions as a recurring, reputationally weighted scorecard on whether frontier labs are operationalizing the safety commitments they have publicly made. The Winter 2025 edition is paired with a Summer 2025 edition earlier in the year, and the recurring schedule (twice yearly) is now established.