Common Elements of Frontier AI Safety Policies (December 2025 Update)

Frontier AI safety policies are documents in which developers of the largest models say how they will test those models for severe risks and what they will do if the risks appear. METR maintains a side-by-side comparison of them, and this is its December 2025 refresh (metr.org). Companies began publishing these policies voluntarily in September 2023; in May 2024 sixteen companies pledged to do so under the Frontier AI Safety Commitments at the AI Seoul Summit, and four more have joined since. Twelve have now published one: Anthropic, OpenAI, Google DeepMind, Magic, Naver, Meta, G42, Cohere, Microsoft, Amazon, xAI and Nvidia.

The useful finding is a shared structure, though not every policy has every piece. METR's original analysis of the first three policies found that each set capability thresholds, levels at which a model could meaningfully help with biological weapons or cyberattacks, or could replicate itself or automate AI research, and committed to evaluations that check whether a model is nearing them, designed to draw out the model's full abilities and run at set times, typically during training, before deployment and after. When a threshold is approached, those policies call for stronger protection of model weights against increasingly sophisticated thieves and for deployment safeguards against misuse, set conditions for halting development or deployment if the mitigations are not enough, point toward accountability measures such as third-party or board oversight, and expect to be revised over time. Across all twelve policies in the current interactive report, deployment safeguards, accountability and plans for updating appear in every one and weight security in eleven, but capability thresholds appear in nine, conditions for halting development in eight, and commitments to elicit full capabilities in seven. METR also stresses that each policy remains distinct; NVIDIA's and Cohere's, for example, emphasize domain-specific risks rather than only catastrophic ones.

The first version, in August 2024, covered only Anthropic, OpenAI and Google DeepMind; a March 2025 edition expanded to twelve policies. This update adds references to some developers' revised policies, without listing which, and brings in two pieces of regulation: the EU AI Act's General-Purpose AI Code of Practice and California's Senate Bill 53. In the interactive version of the report, excerpts from company policies under each element are shown alongside the relevant regulatory guidance, so a reader can see where voluntary commitments and legal expectations line up.