Frontier AI Safety Commitments: who published a framework
At the AI Seoul Summit in May 2024, 20 AI companies (16 at the summit, 4 added in February 2025) signed the Frontier AI Safety Commitments. Each promised to publish a safety framework setting out how it will handle severe risks from its most capable models, by the AI Action Summit in Paris in February 2025. These are voluntary pledges, not law.
| Company | Framework | Status |
|---|---|---|
Amazon United States · signed 21 May 2024 | Amazon's Frontier Model Safety Framework published on February 9, 2025, and updated September 17, 2026 Our analysis | Published |
Anthropic United States · signed 21 May 2024 | Responsible Scaling Policy, Version 3.4 Effective July 8, 2026 Our analysis | Published |
Cohere Canada · signed 21 May 2024 | The Cohere Secure AI Frontier Model Framework V1.0 February 11, 2025 Our analysis | Published |
G42 United Arab Emirates · signed 21 May 2024 | G42's Frontier AI Safety Framework February 2025 Our analysis | Published |
Google United States · signed 21 May 2024 | Frontier Safety Framework, Version 3.1 Published: April 17, 2026 Our analysis | Published |
Magic United States · signed 7 Feb 2025 | AGI Readiness Policy Version 1.0 — July 2, 2024 The page itself now says "This policy is outdated". | Published |
Meta United States · signed 21 May 2024 | Advanced AI Scaling Framework, Version 2 April 7, 2026 (change log, Appendix II; no cover date) Our analysis | Published |
Microsoft United States · signed 21 May 2024 | Frontier Governance Framework February 2026 Our analysis | Published |
Naver South Korea · signed 21 May 2024 | NAVER ASF 2.0: AI Safety Framework 2.0 July 7, 2026 Our analysisVersion 2.0 replaces the August 2024 framework. Official English version published alongside the Korean. | Published |
NVIDIA United States · signed 7 Feb 2025 | Frontier AI Risk Assessment Applicable from August 2025 States that NVIDIA is not currently developing frontier models. | Published |
OpenAI United States · signed 21 May 2024 | Preparedness Framework, Version 2 Last updated: 15th April, 2025 Our analysisAlso published a "Frontier Governance Framework" (2026) for EU and California compliance; the Preparedness Framework remains its safety framework. | Published |
xAI United States · signed 21 May 2024 | xAI Frontier Artificial Intelligence Framework Effective Date: 30 June 2026 Our analysis | Published |
IBM United States · signed 21 May 2024 | Trustworthy AI at scale: IBM's AI Safety and Governance Framework February 7, 2025 Published a governance framework (blog post, Feb 7, 2025) that describes its AI ethics programme but sets no severe-risk thresholds. | Published, not a severe-risk framework |
Samsung Electronics South Korea · signed 21 May 2024 | AI Safety Framework Undated Our analysisCovers on-device consumer AI (privacy, fairness, model governance), not severe or frontier risks. | Published, not a severe-risk framework |
01.ai China · signed 7 Feb 2025 | — No framework found on its site. | Nothing published |
Inflection AI United States · signed 21 May 2024 | — Only a consumer crisis protocol for its Pi assistant. | Nothing published |
Minimax China · signed 7 Feb 2025 | — No framework found on its site. | Nothing published |
Mistral AI France · signed 21 May 2024 | — No framework found. Our Mistral record is an industrial-policy whitepaper, not a safety framework. | Nothing published |
Technology Innovation Institute United Arab Emirates · signed 21 May 2024 | — No framework found on its site. | Nothing published |
Zhipu.ai China · signed 21 May 2024 | — Nothing of its own. Co-authored a general open-weight risk framework with Concordia AI (Sep 2026), hosted on concordia-ai.com. | Nothing published |
Major developers that have published no safety framework
None of these signed the Seoul commitments, and none has published a framework for severe risks from its models. Some signed China’s own industry AI safety commitments; those are not listed here.
- Alibaba (Qwen) · China
Product-security material and the Qwen Guard safety model only; no risk framework.
- Baidu · China
Contributed to China's national AI Safety Governance Framework, a government document, not a framework of its own.
- DeepSeek · China
Publishes model cards in a transparency centre; no risk framework.
- ByteDance (Seed) · China
Security tooling only; no risk framework.
- Moonshot AI · China
The Kimi K2 model card carried safety evaluations; later releases did not. No risk framework.
- Tencent · China
Its research institute publishes governance surveys, not a company framework.
Signatory list: Frontier AI Safety Commitments, AI Seoul Summit 2024 (UK government; first published 21 May 2024, updated 7 Feb 2025). Every framework was opened on the company’s own website; every “nothing published” was checked against the company’s site and the METR index of frontier safety policies. Checked 30 Sept 2026.