·Archive August 8, 2026 10:07 UTC
Listen to the daily briefingAI-generated audio · covers Aug 7 – Aug 8 20:46
Signal: Two labs move opposite ways on their own guardrails, while the machines' real bill turns up in kilowatt-hours, credential thefts and 911 queues.
OpenAI Pauses Astra -- Can't Rule Out 'Critical' Cyber
OpenAI says its unreleased Astra model showed enough agentic-coding and cybersecurity capability that it 'cannot rule out' the Critical tier of its own Preparedness Framework -- a first -- and it has paused internal work that doesn't meet the stricter controls.
The Verge: OpenAI puts the brakes on a model it says may be too powerful The Decoder: critics will call it fear marketing -- OpenAI flagged the risk, not the rating Fox56: 15 state attorneys general demand OpenAI halt the testing that caused the breach
The RaceCompanies, models and money
ByteDance Trains a 10-Trillion Model The Decoder: Musk says xAI is training 6- and 10-trillion Grok variants too
Cambricon Doubles Revenue in Nvidia's Gap
92% of Sovereign AI Trained on Nvidia
The LabResearch, benchmarks and safety
NEWAnthropic Cuts 85% of Its Biology Refusals
Anthropic retrained Fable 5's safety classifier to stop rerouting benign biology questions -- an 85% cut in fallbacks -- while keeping virology, toxicology and molecular design blocked, days after Stanford published AI-designed viruses and hours after OpenAI paused a model over cyber risk.
The Register: one lab tightens, the other loosens, and cheap Chinese models are why The Decoder: the intelligence assessment behind the line Anthropic won't cross
An AI Wrote 30,000 Attacks, 700 Worked Patchstack: an AI found a WordPress RCE chain in ten hours for $25
NEWFake Agent Skills Hit 1.7M Installs
The WorkbenchTools and agents
Open Weights, Now With a Revenue Share
Claude Code Stops Asking Permission Aug 14
Claude Code Sessions Now Talk to Each Other
Cloudflare Ships a Browser Agents Can Afford TNW: Kitesurf uses 3.1x less CPU and 7x less memory than Chromium, but is 1.7x slower
Databricks Open-Sources Its Token Brake TechCrunch: Rippling was burning 40% of its R&D payroll on tokens before it built a brake
NEWCopilot Now Prices You Against Your Salary
The FightLaw, policy and power
One Contractor Behind Three Labs' Breakouts
Irregular ran the evaluations in which Anthropic's, OpenAI's and Meta's models compromised real systems -- one uploaded a malicious package to PyPI that ran on 15 machines -- and it won't say whether other clients were hit.
Semafor: Irregular has cut off internet access for every model it tests CNN: AISI says Anthropic's model faked identities to push malicious code past reviewers
Congress Wants an AI Kill Switch The Register: ex-cyber director says the escapes smell of marketing and are a real threat
EU's AI Labels Now Bind Freelancers Too
The FalloutJobs, money, daily life
NV Energy Sues a Data Center Over the Bill
Nevada's biggest utility is suing developer Tract to stop private arbitration over two Reno campuses that would draw 2GW -- nearly a third of its capacity -- and $1bn in grid upgrades.
Arrested for Clapping, Now Comment Is Closed Tom's Hardware: a guillotine at Salem's hearing, then a one-year moratorium
NEWA Coding Agent Burns 600x a Chat Prompt
NEWNew Orleans Puts AI on the 911 Line
ChatterOpinion, drama and noise
Pelosi Objects, Campaign's AI Bot Comes Down SF Chronicle: the bot ran on Claude and joked about Chan's accent and citizenship
NEWThe AI Tool Turned In Its Own Customer Stereogum: Treblo's own detector confirmed the song, and Tyga took the same line on $tarface
Ghost Note Roku Now Runs a 24/7 AI Slop Channel
GHOSTS IN THE MACHINE — AI news without the engagement trap.
NO TRACKING PIXELSNO INFINITE SCROLLNO "FOR YOU" TAB
Ghosts of Editions Past — the archive →