CAISI frontier-model evaluations
Issuer: NIST: Center for AI Standards and Innovation
Date: JUN 2025
Status: VOLUNTARY / SOFT LAW in United States, as of 29 JUL 2026
Voluntary pre-deployment testing agreements (Anthropic, OpenAI, Google DeepMind, Microsoft, xAI); EO 14409 cybersecurity testing ⚠ single-source.
The renamed US AI Safety Institute (renamed CAISI, 3 Jun 2025) runs voluntary pre-deployment evaluations of frontier models: agreements with Anthropic, OpenAI, Google DeepMind, Microsoft and xAI by May 2026, focused on cyber/bio/national-security risk. Follow-on EO 14409 (2 Jun 2026) reportedly formalized voluntary frontier-model cybersecurity testing via CAISI.
EO 14409 (June 2026 CAISI cybersecurity EO): verified from one secondary source only.
What it actually requires (3 provisions)
- Voluntary pre-deployment evaluations of frontier models (cyber/bio/national-security focus)
- Agreements with Anthropic, OpenAI, Google DeepMind, Microsoft, xAI by May 2026
- EO 14409 (2 Jun 2026): voluntary frontier-model cybersecurity testing ⚠ single-source
How its status moved
- RENAMEDFrom US AISI to CAISI, 3 Jun 2025
- AGREEMENTSFive frontier developers by May 2026
- CURRENTActive: voluntary testing only
Sources (1)
EO 14409: Frontier-model cybersecurity testing EO
Reportedly formalized voluntary frontier-model cybersecurity testing via CAISI (2 Jun 2026), per policy trackers.
EO 14409 (June 2026 CAISI cybersecurity EO): verified from one secondary source only.
The rest of the United States stack
15 more instruments in this jurisdiction, each with its own status, provisions and sources.
- EO 14110 “Safe, Secure, Trustworthy AI”REVOKEDBiden safety EO; reporting mandates died with it.
- EO 14179 “Removing Barriers…”IN FORCEPro-innovation directive; ordered the AI Action Plan.
- America's AI Action Plan (“Winning the Race”)IN FORCE90+ actions, three pillars; policy, not law.
- EO 14365 “National Policy Framework”CONTESTEDDOJ AI Litigation Task Force (est. 9 Jan 2026) + BEAD funding leverage against state AI laws; intervened against Colorado Apr 2026.
- State-law moratoriumFAILEDStripped 1 Jul 2025, Senate 99 to 1.
- NIST AI RMF 1.0 + GenAI ProfileVOLUNTARY / SOFT LAWGovern/Map/Measure/Manage; safe harbor in TX TRAIGA; embedded in procurement.
- OMB M-25-21 & M-25-22IN FORCEAgency AI use & acquisition rules; bind federal agencies only.
- TAKE IT DOWN Act (Pub. L. 119-12)IN FORCEFirst federal AI-content law; FTC 48-hr takedown enforcement since 19 May 2026; up to $53,088/violation.
- Colorado SB 24-205 (Colorado AI Act)REPEALEDFirst comprehensive state AI law; stayed 27 Apr 2026 after xAI suit + DOJ intervention; never took effect.
- Texas TRAIGA (HB 149)IN FORCEProhibited-practices model; AG enforcement; $10k to $200k/violation; NIST safe harbor; sandbox.
- California cluster: SB 53 · AB 2013 · SB 243 · SB 942IN FORCESB 53 frontier transparency (>10²⁶ FLOPs + $500M revenue; OES incident reporting), AB 2013 training-data summaries, SB 243 companion chatbots (PRA ≥$1,000), SB 942 watermarking (UPCOMING 2 Aug 2026).
- Utah AI Policy ActIN FORCEDisclosure-on-request; no 'AI did it' defense; Learning Laboratory; ≤$2,500/violation.
- Illinois HB 3773 + AIVIA (2020)IN FORCEAI employment discrimination = civil-rights violation; first US AI-hiring law for video interviews.
- NYC Local Law 144IN FORCEFirst US bias-audit mandate; $500 to $1,500/day; Dec 2025 Comptroller audit found enforcement 'ineffective'.
- Sectoral enforcers: FTC · FDA · EEOC · CFPBIN FORCEFTC (Operation AI Comply; Growth Cave $48.6M), FDA (1,247+ AI devices; PCCP), EEOC (guidance pulled, statutes intact), CFPB (ECOA/Reg B; Apr 2026 rule narrows disparate impact ⚠).
Where this sits in the wider picture
- The United States regime dossier gives the doctrine this instrument belongs to, next to the other two jurisdictions.
- The Framework Explorer, filtered to US lists every instrument in this jurisdiction in one filterable index.
- The governance simulator shows what these rules do to a real AI system, next to what the other two jurisdictions do to the same one.
- The timeline places this date beside what the other capitals were doing that month.
Knowing the instrument is step one. Complying with it is the job.
The programs teach the work that follows a rule like this one: classification calls, conformity assessment, filings, documentation, and the judgment to defend every decision.
VERIFIED 29 JUL 2026. Every fact on this page is drawn from the sources listed above and dated to the day it was checked. Study aid, not legal advice: the official texts are always authoritative.