Common Elements of Frontier AI Safety Policies (December 2025 Update)
Source
metr.org
Date
Why it matters
Compares what twelve labs commit to on capability thresholds, security and deployment safeguards, plus new EU and California requirements. Useful for engineers anticipating the gating of models they build on.
Key takeaways · AI-distilled
Sixteen companies pledged frontier safety policies at the May 2024 AI Seoul Summit, with four more since. Twelve have published: Anthropic, OpenAI, Google DeepMind, Magic, Naver, Meta, G42, Cohere, Microsoft, Amazon, xAI and Nvidia.
The policies share a structure: capability thresholds for risks such as bioweapons, cyberattacks, autonomous replication or automated AI R&D, evaluations to detect approaching them, and weight security plus deployment safeguards in response.
The policies set conditions for halting development and deployment if mitigations are insufficient, and typically time evaluations before deployment, during training and after deployment.
