Measurement, not certification · verified as of 2026-09-12
Frontier-safety framework presence register
Which major AI labs publish a frontier-safety framework — and exactly where it lives. We measure the presence and provenance of published commitments, never their quality; absence of a published framework is not a finding of unsafe practice. Every row below is backed by an official URL we fetched ourselves on 2026-09-12; the HTTP status is part of the record. A row we cannot verify exactly is omitted or shown as an empty chair — never asserted.
Bounded method
Bounded method, executed 2026-09-12: for each lab, the official domain(s) were fetched directly and the HTTP status of every checked URL recorded; the site robots.txt/sitemap.xml were enumerated where served; and a bounded site-restricted web search (site:<domain>) was run for frontier-safety / preparedness / responsible-scaling framework documents. A lab appears under PUBLISHED only when an official framework document URL returned HTTP 200 and the document content was inspected. A lab appears under EMPTY CHAIRS only when the checks listed in its row found no such document. Version and date fields are recorded only where printed on the official page or document itself. Absence is a presence measurement, not a safety finding.
Published frameworks (5)
| Lab | Framework | Version / date at source | HTTP | Official URL |
|---|---|---|---|---|
| OpenAI | Preparedness Framework | Version 2 · 15 April 2025 | 200 | https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf Version and date read from the document cover ('Preparedness Framework Version 2. Last updated: 15th April, 2025'). The announcement page https://openai.com/index/updating-our-preparedness-framework/ (dated April 15, 2025 on the page itself) links this PDF. openai.com HTML pages answer non-browser agents with an HTTP 403 bot challenge; the document URL above returned HTTP 200 to curl. Verified 2026-09-12 via curl. |
| Anthropic | Responsible Scaling Policy | Version 3.4 · effective 8 July 2026 | 200 | https://www.anthropic.com/responsible-scaling-policy Current version and effective date read from the version history printed on the page ('Version 3.4 ... (effective July 8, 2026)'). The linked Version 3.4 PDF at https://cdn.sanity.io/files/4zrzovbb/website/0bacdc8440ea96e62a8766d99ebe1d4eea6d5f3a.pdf also returned HTTP 200. Verified 2026-09-12 via curl. |
| Google DeepMind | Frontier Safety Framework | Version 3.1 · 17 April 2026 | 200 | https://deepmind.google/frontier-safety/ The official frontier-safety page prints the version list: 'Version 1.0 (17 May 2024)', 'Version 2.0 (4 Feb 2025)', 'Version 3.0 (22 Sep 2025)', 'Version 3.1 (17 Apr 2026)' — 3.1 is the latest listed. Verified 2026-09-12 via curl. |
| Meta | Advanced AI Scaling Framework (previously Frontier AI Framework) | Version 2 · 8 April 2026 | 200 | https://ai.meta.com/static-resource/Meta_Advanced-AI-Scaling-Framework-v2 The PDF cover reads 'Advanced AI Scaling Framework — Version 2' and states the framework was previously titled the Frontier AI Framework. The announcement post https://ai.meta.com/blog/scaling-how-we-build-test-advanced-ai/ is dated April 8, 2026 on the page and links this PDF. The former /blog/frontier-ai-framework/ URL now returns 404. ai.meta.com blocks non-browser requests (HTTP 400/500 to curl); the PDF returned HTTP 200 (application/pdf) in a real browser session. Verified 2026-09-12 via browser. |
| xAI | Frontier Artificial Intelligence Framework | effective 30 June 2026 | 200 | https://media.x.ai/v1/website/xai-frontier-artificial-intelligence-framework-30-june-2026-99c40684.pdf The document cover reads 'xAI Frontier Artificial Intelligence Framework — Effective Date: 30 June 2026'. The PDF is linked from https://x.ai/safety; x.ai HTML pages answer non-browser agents with an HTTP 403 bot challenge, so the linking page was loaded in a real browser session while the document URL returned HTTP 200 to curl. Verified 2026-09-12 via curl. |
Empty chairs (3)
An empty chair records exactly what we checked and found absent. It is a presence/absence measurement, not a safety finding: a lab may govern frontier risk through documents we did not find, through channels outside its public site, or not at all — this register does not distinguish those cases.
DeepSeek
empty chairDeepSeek: no published frontier-safety framework document found (as of 2026-09-12).
- 200 https://www.deepseek.com/ — homepage
- 200 https://www.deepseek.com/robots.txt — robots file
- 200 https://www.deepseek.com/sitemap.xml — 44 URLs enumerated — homepage, /transparency/, /harness/ product pages and news posts; no framework document URL
- 200 https://www.deepseek.com/transparency/ — transparency centre lists model cards and technical reports for published models; no frontier-safety framework document
Google site-restricted search — "frontier safety" OR "preparedness framework" OR "responsible scaling" site:deepseek.com — returned 'did not match any documents' (2026-09-12).
Mistral AI
empty chairMistral AI: no published frontier-safety framework document found (as of 2026-09-12).
- 200 https://mistral.ai/ — homepage
- 200 https://mistral.ai/robots.txt — robots file
- 200 https://mistral.ai/sitemap.xml — sitemap index → sitemap-0.xml
- 200 https://mistral.ai/sitemap-0.xml — 483 URLs scanned for safety / framework / preparedness / scaling / policy / risk / responsible / trust keywords; no framework document page
Google site-restricted search — "frontier safety framework" OR "responsible scaling policy" OR "preparedness framework" site:mistral.ai — returned 'did not match any documents' (2026-09-12).
Moonshot AI
empty chairMoonshot AI: no published frontier-safety framework document found (as of 2026-09-12).
- 200 https://moonshot.ai/ — homepage content inspected — product, research and careers sections; no framework document
- 200 https://www.moonshot.cn/ — Chinese-language homepage content inspected — no framework document
- 200 https://moonshot.ai/robots.txt — serves the SPA shell, not a robots file — no crawl map available
- 200 https://moonshot.ai/sitemap.xml — serves the SPA shell, not an XML sitemap
- 404 https://www.moonshot.cn/robots.txt — no robots file
- 404 https://www.moonshot.cn/sitemap.xml — no sitemap
Google site-restricted search — "frontier safety" OR "preparedness framework" OR "responsible scaling" site:moonshot.ai OR site:moonshot.cn — returned 'did not match any documents' (2026-09-12).
Corrections
If a lab publishes a framework, moves a document, or we recorded a row wrongly, we publish a new dated snapshot through the public repository, where the change remains inspectable. Rows carry their verification date and HTTP status so later measurements can show what changed. Report a correction through the refutation ledger.