AI Safety and Regulation Turmoil: From Anthropic Fable Incident to OpenAI Industrial Policy

AI 안전 | Fri Jun 19 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 6 sources

Current AI safety and regulatory issues include Anthropic Fable model export controls, OpenAI's industrial policy proposal, the UK facial age estimation controversy, and Claude robot autonomous control experiments.

Analysis

[Trump White House & Anthropic] imposed export controls on Fable 5 and Mythos models ^[2][3][4]

• Executive order blocking foreign access issued on Friday evening
• Anthropic took both models entirely offline due to compliance difficulties
• Access blocked for Fortune 500 customers including Apple and Meta
• as well as foreign national employees
• Triggered by jailbreak concerns raised by Big Tech executives including Andy Jassy

[Wired Analysis] diagnosed the 'Wild West' phase of US AI regulation ^[4]

• Almost no written laws governing frontier AI
• Trump repealed Biden's AI executive order and is pushing to nullify state AI laws
• Opacity from White House not clearly publicizing Anthropic's violations
• David Sacks's X post serves as effectively the only official position

[OpenAI] proposed Industrial Policy for the Intelligence Age ^[1]

• Released people-first policy ideas to prepare for the superintelligence era
• Pilot program offering up to $100
• 000 research grants and $1M API credits
• Opened OpenAI Workshop in Washington DC in May
• Reviewing grants after receiving over 400 pieces of feedback

[UK Home Office] pushed forward facial age estimation for asylum seekers ^[5]

• AI facial age estimation system to be applied at UK borders starting next year
• Internal testing confirmed average error of 4.6 years for Sub-Saharan African women
• Risk of 13.5-year-old girls being classified as 18-year-old adults
• Weakened governance including disbanding of the advisory science committee

[Anthropic Frontier Red Team] evaluated Claude's autonomous robot control capability with Project Fetch Phase Two ^[6]

• Claude Opus 4.7 performed robodog control tasks without human assistance
• Achieved approximately 20x faster speed than the fastest human team a year ago
• However
• still failed precise manipulation tasks such as fetching a beach ball
• The 'model assists human → human assists model → model performs autonomously' pattern seen in cybersecurity is extending to the physical world

Sources

• [1] Industrial policy for the Intelligence Age - OpenAI Blog
• [2] Who decides when AI is too dangerous? - The Verge AI
• [3] Vibe-decoding the White House-Anthropic fight over Fable - The Verge AI
• [4] The White House Is Making Up Its Rules for AI in Real Time - Wired AI
• [5] The UK Will Scan Asylum-Seekers’ Faces for Age Checks—Despite Knowing the Tech Is Flawed - Wired AI
• [6] Frontier Red TeamProject Fetch: Phase two - Anthropic Research