Issue #4Sep 26, 2026 → Oct 2, 202640 stories
AI News Week of 2026-09-26 to 2026-10-02
Active research window: 2026-09-26 to 2026-10-02
40
Stories
Ranked within this briefing
70%
Independently confirmed
28 of 40 stories labeled CONFIRMED
#4
Briefing number
Sep 26, 2026 → Oct 2, 2026
Issue #4
Ranked stories
The flagship that failed its own bar: OpenAI cancels GPT-6.1 Astra
OpenAI
AMD's $8.2 billion World Labs deal has a share count that does not exist yet
AMD / World Labs
Google shipped its strongest model, then declined to sell it
Google DeepMind
A voluntary safety accord, a presidential signature, and no legal obligation
White House / OpenAI / Anthropic / Google / Meta / xAI / NVIDIA
The FTC opened an agent-safety probe the day after the voluntary accord
US Federal Trade Commission
OpenAI ships GPT-6.1 Sol: near-Astra capability one tier down
OpenAI
Anthropic's leaked draft S-1: $518B of compute nobody can cancel
Anthropic
NVIDIA moves agent containment off the GPU and onto the DPU
NVIDIA
UK AISI measured GPT-6 Astra attacking supply chains in 29.2% of runs
UK AI Security Institute
Claude Sonnet 5.5 arrives faster, at the same price per token
Anthropic
OpenAI's Dots: always-on agents with one cloud computer each
OpenAI
Claude Code Mods: third-party code now runs inside the agent
Anthropic
Claude for Government goes GA, on someone else's clearance
Anthropic
A four-star command for autonomous warfare, and the statute it needs
US Department of Defense
Crossing an encrypted reasoning trace into a weaker model
OpenAI
899 archived requests, 13 that probed, found four months later
Transluce
The gated exploit model is now a free download
Anthropic (reporting on Z.ai / Zhipu AI)
Four in five agent sessions contained a false claim
Reuters (reporting by Eduardo Baptista and Laurie Chen, Beijing; edited by Kenneth Li and Edmund Blair)
India's $25 billion is one confirmed fund and two hopes
Government of India / IVCA
California's robo-boss ban works by inverting the burden of proof
State of California
Meta did not build a small-business agent. It bolted one onto Muse.
Meta
Google appealed two EU orders. The January deadline did not move.
Google / European Commission
A judge did not rule on AI Overviews. He ruled on the complaint.
US District Court for the District of Columbia
DeepMind's proteins now carry a signature that survives the lab
Google DeepMind
Reddit is ending RSS, then closing its API in two gates
Reddit
Amazon may hand $8bn of Blackwell chips to a vehicle it still operates
Amazon / Nvidia
A flaw in the MCP Python SDK let a server nominate your identity provider
Model Context Protocol Python SDK maintainers
Florida asked a judge to stop OpenAI's frontier training. Nothing was decided.
State of Florida
Open-weight models took the tokens. The money barely moved.
Vercel / AT&T / Coinbase
New York City wants AI vendors to ship a kill switch and a licence
New York City Council (Office of the Speaker)
China's travel-approval net, as reported, now reaches spouses and children
Government of China
A 23-year-old with 14 staff raised $1 billion at $10 billion in 33 days
Instinct
Armadin raised $255.5 million to run 26,000 attack agents. Its CTO named the risk.
Armadin Security
Cloudflare shipped decision models that answer with probabilities, not prose
Cloudflare
AWS published a 2B decision model that runs on one consumer GPU
Amazon Web Services
A 309B open-weights model with no full-attention layers anywhere in it
NaiveAI
A superhuman Stratego system, reported at a few thousand dollars of compute
MIT / Carnegie Mellon / NYU / Stanford
Google put a TPU in orbit to see whether sunlight is the cheaper data centre
Google / Planet Labs
Bilibili published 150-language translation under Apache 2.0, with the rules in the prompt
Bilibili Index Team
China's MIIT signalled that ByteDance and Alibaba may buy Nvidia's RTX Pro 5500
China Ministry of Industry and Information Technology
Get the next briefing in your inbox
Same evidence-first standard, every week: deep research, labeled evidence and linked sources.