Center for AI Standards and Innovation (CAISI)

CAISI is the US federal body that sets AI standards and evaluates frontier-model safety — the government’s in-house technical evaluator for AI. In the 20-day Anthropic showdown it was CAISI and the NSA that judged Anthropic’s first round of fixes insufficient, forcing further changes before the Fable 5 / Mythos 5 models were cleared for re-release on 1 July 2026 anthropic-models-revived-2026.

Role on the wiki’s map

CAISI is the concrete answer to a question the export-control thread kept raising: who actually adjudicates whether a frontier model is safe to ship? Under EO 14409‘s pre-release submission rule, CAISI (with the NSA) is the technical gate that says “good enough” or “not yet.” That makes it the US operational counterpart to the assurance role played voluntarily by iso-iec-42001 certification and the nist-ai-rmf — but wielded as a binding release gate on national-security grounds, not a voluntary standard. It is the standards-institute-as-regulator node the spoke’s inward-pointed export-control instrument runs through.

CAISI is the successor to the US AI Safety Institute — the same “safety → security/standards” renaming the UK made when its AI Safety Institute became the AI Security Institute (Feb 2025), which the synthesis tracks as soft-oversight drifting toward a harder security posture.

Caveat

CAISI’s role here is reported via anonymous US officials in a single Axios account; the fact that a federal standards body plus the NSA gated the release is the load-bearing claim, the internal deliberation detail is secondhand.

Related

anthropic-models-revived-2026 · eo-14409 · export-controls-on-ai · howard-lutnick · nist-ai-rmf · iso-iec-42001 · ai-governance · synthesis