AI Risk Tiers

How laws and labs sort AI systems into risk levels.

The EU pushed its high-risk AI obligations back to December 2027 and August 2028 after Council approval on June 29, 2026. China tiers agents by who holds the decision; California and New York tier frontier models by compute above 10^26 operations.

Critical incident reporting deadlines: California SB 53 15 days; New York RAISE 72 hours; Either, imminent harm 24 hours CRITICAL INCIDENT REPORTING DEADLINES California SB 53 15 days New York RAISE 72 hours Either, imminent harm 24 hours HOURS
Frontier compute threshold
10^26 ops
RAISE repeat-violation penalty
$3M
China agent decision tiers
3
EU risk tiers
4

Evidence

Labs moved away from hard thresholds as legislators adopted them

Documented cases: AI Risk Tiers
DateCaseCountryWhat happenedSource
May 2025Anthropic activates ASL-3United StatesFirst activation of the ASL-3 safeguard tierSource
Sep 29, 2025SB 53 signedUnited StatesFrontier frameworks, transparency reports, whistleblower protectionsSource
Feb 24, 2026Anthropic RSP v3United StatesReplaces hard thresholds with a public roadmap and periodic risk reportsSource
Mar 27, 2026RAISE Act amendmentUnited StatesReshaped to mirror SB 53, with 72-hour reportingSource
May 8, 2026China agent Implementation OpinionsChinaThree decision tiers; filing and recall in four sectorsSource

Rules

Three tiering logics: by use (EU), by compute (US states) and by decision authority (China)

Rules that address risk tiering
RuleJurisdictionWhat it requiresStatus
EU AI Act tiersEUUnacceptable, high, limited and minimal riskHigh-risk delayed to 2027 and 2028
California SB 53CaliforniaFrontier frameworks and incident reportingIn force Jan 2026
New York RAISE ActNew YorkFrontier frameworks; 72-hour incident reportingEffective Jan 1, 2027
China agent Implementation OpinionsChinaThree decision tiers, filing and recallEffective Jul 15, 2026
OpenAI Preparedness Framework v2CompanyHigh and Critical capability thresholdsApr 2025

Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.

Record

Page Record

Risk group
Risks Hub
Datasets
ai_risk_evidence.csv (5 rows); registry incidents (0 linked); ai_attack_surface.csv
Key sources
Anthropic activates ASL-3
SB 53 signed
Anthropic RSP v3
RAISE Act amendment
China agent Implementation Opinions
Last verified
October 2026