METR, 2026-01-29; retrieved 2026-09-14 tier 3 self (metr)confirmed

Frontier AI safety regulations: A reference for lab staff

https://metr.org/notes/2026-01-29-frontier-ai-safety-regulations/

EU CoP: adequate access incl. helpful-only versions, 20 business days, external evaluators each new frontier model; enforceable Aug 2026. Cited from a search snippet; not re-fetched in full. [verify 2026-09-15 grok-pass] SB 315: independent auditor, no financial interest, pay not tied to findings; audits from 2028.

Signals citing this source

EU AI Office, A (for): Legal power to compel; the only body whose access does not depend on goodwill.
EU AI Office, S (for): Code sets the access floor: helpful-only versions where security allows, at least 20 business days, external evaluators for each new frontier model.“A developer should engage independent external evaluators for each new frontier model”
EU AI Office, R (against): Safety and Security Model Reports are submitted to the Office, not the public.
EU AI Office, F (for): Public funding.
EU AI Office, G (for): Regulator with statutory basis.
EU AI Office, P (for): Civil-service conflict rules.
EU AI Office, X (for): No commercial products.
EU AI Office, S (for): The AI Office holds the legal power under the AI Act to evaluate general-purpose models with systemic risk and to compel access and information; scope for those evaluations is set by the regulator, not the developer.

Ledger rows citing this URL

none