Safety and policy in AI

Model safety, usage policies, trust and safety, regulation. Counted from the record's checked claims, one column per quarter since 2023. /trends/safety-policy.md

2023 Q12026 Q4*

By quarter

2026 Q4*	9 of 155 updates (6%)
2026 Q3	15 of 279 updates (5%)
2026 Q2	15 of 233 updates (6%)
2026 Q1	2 of 120 updates (2%)
2025 Q4	2 of 89 updates (2%)
2025 Q2	4 of 44 updates (9%)
2025 Q1	1 of 32 updates (3%)
2024 Q2	1 of 28 updates (4%)

all quarters 2024 Q2 2025 Q1 2025 Q2 2025 Q4 2026 Q1 2026 Q2 2026 Q3 2026 Q4

1–9 of 9 AI claims in 2026 Q4, newest first

2026-10-29valid

OpenAI · policy · Updated Model Spec

clm_FFC2E2E785A0 · valid · 1 source (1 primary) · conf 0.70

2026-10-28valid

Google · released · Gemini 4 Carbon

clm_0516802A12DE · valid · 1 source (1 primary) · conf 0.70

2026-10-27valid

Microsoft · model release · MAI Cyber 1.1 Flash

clm_61E141829B00 · valid · 1 source (1 primary) · conf 0.70

2026-10-22valid

Cohere · launched · Confidential Computing on Cohere Model Vault

clm_FB95B6CC521B · valid · 1 source (1 primary) · conf 0.70

2026-10-20valid

Claude · feature · Web fetch domain restrictions

clm_FBFD57ACC84D · valid · 1 source (1 primary) · conf 0.70

2026-10-12valid

AI Security Institute · policy · national AI risk defences

clm_885C98452084 · valid · 1 source (1 primary) · conf 0.70

2026-10-08valid

Anthropic · launched · Cyber Mission

clm_0C637C075926 · valid · 1 source (1 primary) · conf 0.70

2026-10-08valid

Anthropic · policy · New Usage Policy version

clm_C1868F7DE4CA · valid · 1 source (1 primary) · conf 0.70

2026-10-06valid

Anthropic · launched · Cyber Verification Program

clm_977BA9105325 · valid · 1 source (1 primary) · conf 0.70