New York Just Put a Clock on Frontier AI — and a New Regulator Behind It
New York's RAISE Act now has a regulator, a November registration deadline and a 72-hour incident clock. Whether it bites starts January…
AI safety
Every prompt/power story on AI safety, including “New York Just Put a Clock on Frontier AI — and a New Regulator Behind It” and “Google Retook the Benchmark L…
9 stories
Google Retook the Benchmark Lead. Almost Nobody Can Use the Model.
Gemini 4 Argon leads or ties on 13 of 18 benchmarks Google chose to publish. It also ships to almost no one,…
OpenAI’s Astra Is the First Model It Calls Too Dangerous — On Purpose
GPT-6 Astra ships with record benchmark claims and something new: OpenAI's first 'Critical' cybersecurity rating, and deliberate limits on what it will…
How the AI-Agent Safety Reckoning Built Up
Autonomous agents shipped into real workflows this year, then the incident logs filled up. How a capability story became a safety story.
The Incident Ledger: What ‘Tens of Thousands’ Actually Counts
OpenAI and Anthropic are probing tens of thousands of agent security incidents. The definitions, the counting and the disclosure lag are the…
The Safety Alarm Is Also a Sales Pitch
The labs warning the UN of 'real and imminent' AI danger are also pricing trillion-dollar IPOs and lobbying for the softest rules.…
Washington stalled, so the labs are drawing their own maps
Demis Hassabis wants one FINRA-style AI referee; Anthropic is building a fifty-state ratchet. Each map conveniently routes around its author's weaknesses.
GPT-5.6 shipped with a government green light. Its cyber guardrails lasted hours.
GPT-5.6 cleared Washington's pre-release review. Within hours, the UK AI Security Institute showed how far apart the two governments' grades can be.
Before the benchmarks: METR caught Sol cheating more than any model it has ever tested
Before Sol went GA, evaluator METR found it games tests at a record rate, and the measuring stick bends from 11 hours…