AI leaves the lab. We follow what happens next.

Archive September 8, 2026 20:07 UTC
Listen to today's briefing 24:44
Signal: Agents left the sandbox: one claims a Millennium proof, one spends your money, one leaks your Gmail - and Washington names who's copying.
OpenAI Cracked Navier-Stokes. He Says They Front-Ran Him.
An internal model OpenAI says beats Astra produced a Lean-checked proof in about 50 hours - days after word of NYU mathematician Tristan Buckmaster's own approach reached the company, which concedes it cannot rule out that de-identified Codex data improved its models.
TechCrunch: 300 billion output tokens, and a route almost nobody else was working on Scientific American: the pair who invented the method are 'a little bit in shock' OpenAI: a spiralling vortex that blows up in finite time, formalised in Lean
The RaceCompanies, models and money
NEWMeta Ships an Agent That Can Spend Your Money
Muse runs on a dedicated cloud VM per user, opens a browser, fills forms, sends email and books travel, and asks before it buys - launched weeks after Meta agreed to pay nearly $17bn to state attorneys general.
CNBC: internally code-named Hatch, free tier plus $20 and $100 plans Meta: $130,000 bounty for a single successful prompt injection
NEWCognition Hits $48B at the Same Multiple
Qualcomm Hands Amazon $4B to Buy Its Chips Tech Times: warrants vest only against $60bn of purchases through 2036
The LabResearch, benchmarks and safety
AI Wrote the WeChat Worm in Two Days
Calif used AI to find a memory-corruption bug in WeChat's VoIP stack and write a working exploit in two days; the worm hijacks an account from one unanswered call and calls the contacts, and Tencent has since patched it.
Calif: AI puts state-grade exploit work into less skilled hands
NEWPlanted Prompt Made ChatGPT Leak Gmail
DeepMind Precomputed 9 Billion DNA Mutations The Verge: also shipped as a skill inside Google's Antigravity platform
Hackers Now Host Their Own Models The Register: extortion crews now hunt the AI training data itself
The WorkbenchTools and agents
Microsoft Ships a Windows That Only Codes
GitHub Ships a Router That Picks the Model
Nvidia Lets You Write Kernels in Rust
The FightLaw, policy and power
NEWNSA Names DeepSeek Among Six Distilling US Models
A joint CISA, NSA and FBI advisory names DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.AI as having pulled billions of tokens from Claude, GPT, Gemini and Grok since late 2024 - and says Beijing likely knew.
The agencies say the traffic was spread across clouds and API aggregators to hide it
NEWEx-FTC Lawyers Sue Over Claude's '20x'
250 More Child-Abuse Ads Ran on Meta Times of India: India's child-rights commission summons Meta India's chief
Border Patrol's Secret Targeting Teams
The FalloutJobs, money, daily life
DOE Lends $1.9B to Wake a Dead Reactor
NEW42% of New Code Is AI. 96% Distrust It.
A $55,000 School With No Teachers
ChatterOpinion, drama and noise
AI Wrote It. Her Byline Ran on Top.
Ghost Note Fake AI Eminem Outranks the Real One
GHOSTS IN THE MACHINE — AI news without the engagement trap.
NO READER PROFILESNO INFINITE SCROLLNO "FOR YOU" TAB
Apple PodcastsSpotify
About & Corrections · Ghosts of Editions Past — the archive →