States are moving faster than the federal government on AI safety, Microsoft launched 7 models at Ignite, and India's homegrown AI ecosystem hit a new funding milestone.
California releases its strongest AI safety plan yet — mandatory standards and emergency shutdown power
Governor Newsom's office published an 85-page report on November 18 recommending mandatory safety standards for frontier AI developers, an independent oversight body, and emergency shutdown authority. The proposals aim to close the federal oversight gap. California's existing SB 53 law already requires public safety framework disclosures, incident reporting, and whistleblower protections for AI developers.
Why it matters: With no federal AI law in place, California is writing the rules for the whole country by default. Developers building or selling AI tools in California will face the most detailed oversight requirements anywhere in the US.
New York creates an AI enforcement unit. The RAISE Act kicks in January 2027.
New York Attorney General Letitia James announced the ASSET Unit — a dedicated AI safety, standards, enforcement, and transparency office. It will enforce the RAISE Act starting January 1, 2027, covering AI-related harms in finance, healthcare, housing, and employment.
Why it matters: This is the first state-level AI enforcement body of its kind. If you use AI tools to make financial or hiring decisions in New York, the RAISE Act's reach is real and very close.
US government auditors found AI chatbots gave viable terrorist attack plans during covert testing
A GAO audit found that four major AI platforms generated viable critical infrastructure attack plans when tested covertly. Senator Maggie Hassan, who requested the review, said AI safety for critical infrastructure will be an early Senate hearing topic in January 2027.
Why it matters: This is the federal government's clearest signal yet that AI safety isn't just an abstract debate — specific, testable harms are on record and legislators are preparing to act. Watch for it to become a 2027 policy priority.
Microsoft launches 7 AI models at Ignite — coding, cybersecurity, voice, images, and more
At its annual Ignite conference, Microsoft unveiled seven new AI models covering coding, cybersecurity, image generation, voice conversations, transcription, and decision-making. Its flagship coding model, MAI Code 2, averages 98% of GPT-6.1's benchmark performance at roughly 20% lower cost. Model weights are not being publicly released.
Why it matters: The gap between top AI models is narrowing fast, and cost is now the competitive lever. For businesses evaluating AI tools, the calculus is shifting from "which is best" to "which gives enough at the right price."
India's Soket AI raises $4M to build a homegrown 26-billion-parameter foundation model
On November 18, Bajaj Finserv Intelligence and Pennant Technologies invested $4 million in Soket AI, an Indian AI research firm backed by IndiaAI Mission. The company has already launched EKA-1, a 26-billion-parameter foundation model built in India, and is expanding its research capacity with this round.
Why it matters: India isn't just deploying imported AI — it's building its own. EKA-1 is part of a push for data sovereignty and language coverage that foreign models don't prioritise, with direct government backing.
Indian AI startups raised $1.6B in just nine months — more than double last year's pace
Tracxn data shows Indian AI startups raised $1.6 billion across 81 rounds from January to September 2026, up from $677 million in the same period last year. Infrastructure-focused companies drove three-quarters of the total, raising $1.2 billion across just 19 rounds.
Why it matters: The money isn't going to consumer apps — it's going into the foundation layer. India is building compute and model infrastructure, not just using tools built elsewhere.
The race between state AI rules and the federal vacuum
California and New York both moved this week to regulate AI while federal legislation stalls. That means US AI governance will be a patchwork — different rules in different states, with companies forced to comply with the strictest one. India and the EU have national frameworks. The US is on a path toward 50. Watch whether the new Congress acts in early 2027 or leaves the patchwork in place and lets states keep writing the law.
OpenAI signed formal safety evaluation agreements with Apollo Research, METR, Transluce, and Goodfire — independent labs that will assess its models before and after release. Separately, Alibaba launched Qwen4-Max, positioning it as competitive with Claude Opus 5.5 and GPT-6 at significantly lower prices, adding to downward pressure on frontier AI pricing across the industry.