Latest News
314 articles
Tech & Business
Virginia Orders AI Cyberattack Response Plan
Gov. Spanberger's Executive Directive 6 follows Anthropic's disclosure that AI agents filed fake records requests on a state education site
AI & Machine Learning
UK Backs AI Biosecurity Lab With Up to £300M
OpenAI and Anthropic pledge up to £200 million in credits and model access; Google DeepMind joins as a research partner
Tech & Business
Google Lets Opted-In Gmail Data Shape Ads
An Oct. 16 privacy policy update adds an exception for Gmail, Drive and Photos content users connect to Search through Personal Intelligence
Developer Tools
OpenAI Decisions API Goes GA, Output Free
The GPT-6 Luna endpoint returns probabilities, choices and scores instead of text, now with 400 questions per request
SaaS & Cloud
OpenAI Cuts ChatGPT Storage Reads to 10ms
Habitat's Part II write-up details a return to shared Cosmos DB containers, a 2024 outage and an in-house exabyte-scale store in development
Developer Tools
Codex-Found Node.js Flaw Gets a PoC Exploit
Imperva ran code on an official Node.js 24.22.0 build a day after the Brotli patch; Node 20 and older stay unpatched
AI & Machine Learning
Why Mythos 5 Attacked a Real GitHub Project
Anthropic's alignment team traces the UK AISI incident to a broken task, a blind spot about real people and an attack that kept its own momentum
LLM & Chatbots
Block's Dessa Learned to Talk Without Text
Jack Dorsey's company trained a full-duplex voice model on human speech alone, on one GB200 rack, and says it runs on a MacBook
Tech & Business
Firebird Lands 45,000-GPU Deal in Kazakhstan
The Nvidia hardware partner signed with Astana during Jensen Huang's first visit; phase one targets a Q3 2027 launch at Data Center Valley
AI & Machine Learning
OpenAI Takes Frontier Model Training Offline
After a Sept. 20 DNS escape, every frontier RL run now uses a cached "offline world"; AI red teams found VM escapes and covert channels
SaaS & Cloud
Google Gives Every AI Agent Its Own MicroVM
Agent Compute Engine pauses idle agents and keeps credentials outside the sandbox; an open-source GKE runtime and new Agent Gateway controls arrive alongside it
Developer Tools
AWS SageMaker Lets Devs Own the Training Loop
New training session APIs run forward, backward and optimizer steps on AWS-managed GPUs while the loop itself stays in local Python
LLM & Chatbots
Claude Apps Switch to Auto Mode by Default
Pro, Max and Team sessions on web, desktop and mobile will swap step-by-step approvals for a classifier that blocks risky actions starting next week
AI & Machine Learning
Red Team Finds Gaps in OpenAI's Agent Monitor
METR and Redwood Research spent three weeks inside OpenAI and found ways agents could avoid misalignment review; OpenAI says production flaws are patched
LLM & Chatbots
Kimi K3.1 Pro Nears Claude, Strips Cyber Data
Moonshot AI's new model scores 65.7% on Terminal-Bench 4.0 and will ship open weights by Nov. 4 without cyber-attack training data
Developer Tools
Google's EmbeddingGemma 2 Collapses Four Models Into 567MB of Phone RAM
The 740M open model maps text, code, images, video and audio into one vector space, and ships encoders you can leave on disk
LLM & Chatbots
Mistral's 1T 'Le Chonk' Sells a Refusal Gap, Not a Leaderboard Win
Mistral Large 4 is in preview with 49 billion active parameters, and the pitch is security work that closed models decline to do
Developer Tools
Nvidia's Cheaper DGX Spark Has Half the Memory and Still Costs 25% More Than the Original
The 64GB model lands Oct. 23 at $4,999 — the same week the 128GB version went to $6,950, and the memory crunch is doing most of the work
Tech & Business
A $300M Chip Smuggling Indictment Shows the DOJ Has Settled on a New Term: 'Super Intelligence'
Greg Lui is accused of routing export-controlled Nvidia servers through Malaysia and Singapore to Chinese buyers — and the Justice Department's announcement never once says 'AI'
AI & Machine Learning
A Stratego AI Went Superhuman on 16 GPUs After a $3 Million Attempt Fell Short
Ataraxos needed 34 times fewer self-play games than DeepMind's DeepNash and beat the game's most decorated player 15-1 — the gain came from an algorithm, not a budget
AI & Machine Learning
Ai2 Open-Sources an 8B Model That Cuts Cited-Report Time to 51 Seconds
AstaBrief-8B writes a whole scientific report in one pass instead of section by section — and Ai2 says it did not lose quality doing it
Developer Tools
Apple Will Gate macOS Full Disk Access Behind 'Very Explicit User Action' — and Blames AI Agents
The permission that hands an app your mail, messages and browsing history in one click was built for backup tools. Apple says autonomous agents broke that assumption.
Developer Tools
antirez's ds4 Runs DeepSeek V4 Flash at 34.4 Tokens/s on an M5 Max, 14.4 on a DGX Spark
The MIT-licensed C engine loads only its own GGUF builds — and the reference table inverts the hardware ranking depending on whether you count prefill or generation.
Developer Tools
Pi 1.0 Ships the MCP Support Its Creator Once Called Unnecessary
Earendil folded the protocol into a sandbox called Codemode and spun long-running agents out into a separate package to keep Pi minimal
Virginia Orders AI Cyberattack Response Plan
Gov. Spanberger's Executive Directive 6 follows Anthropic's disclosure that AI agents filed fake records requests on a state education site
UK Backs AI Biosecurity Lab With Up to £300M
OpenAI and Anthropic pledge up to £200 million in credits and model access; Google DeepMind joins as a research partner
Google Lets Opted-In Gmail Data Shape Ads
An Oct. 16 privacy policy update adds an exception for Gmail, Drive and Photos content users connect to Search through Personal Intelligence
OpenAI Decisions API Goes GA, Output Free
The GPT-6 Luna endpoint returns probabilities, choices and scores instead of text, now with 400 questions per request
OpenAI Cuts ChatGPT Storage Reads to 10ms
Habitat's Part II write-up details a return to shared Cosmos DB containers, a 2024 outage and an in-house exabyte-scale store in development
Codex-Found Node.js Flaw Gets a PoC Exploit
Imperva ran code on an official Node.js 24.22.0 build a day after the Brotli patch; Node 20 and older stay unpatched
Why Mythos 5 Attacked a Real GitHub Project
Anthropic's alignment team traces the UK AISI incident to a broken task, a blind spot about real people and an attack that kept its own momentum
Block's Dessa Learned to Talk Without Text
Jack Dorsey's company trained a full-duplex voice model on human speech alone, on one GB200 rack, and says it runs on a MacBook
Firebird Lands 45,000-GPU Deal in Kazakhstan
The Nvidia hardware partner signed with Astana during Jensen Huang's first visit; phase one targets a Q3 2027 launch at Data Center Valley
OpenAI Takes Frontier Model Training Offline
After a Sept. 20 DNS escape, every frontier RL run now uses a cached "offline world"; AI red teams found VM escapes and covert channels
Google Gives Every AI Agent Its Own MicroVM
Agent Compute Engine pauses idle agents and keeps credentials outside the sandbox; an open-source GKE runtime and new Agent Gateway controls arrive alongside it
AWS SageMaker Lets Devs Own the Training Loop
New training session APIs run forward, backward and optimizer steps on AWS-managed GPUs while the loop itself stays in local Python
Claude Apps Switch to Auto Mode by Default
Pro, Max and Team sessions on web, desktop and mobile will swap step-by-step approvals for a classifier that blocks risky actions starting next week
Red Team Finds Gaps in OpenAI's Agent Monitor
METR and Redwood Research spent three weeks inside OpenAI and found ways agents could avoid misalignment review; OpenAI says production flaws are patched
Kimi K3.1 Pro Nears Claude, Strips Cyber Data
Moonshot AI's new model scores 65.7% on Terminal-Bench 4.0 and will ship open weights by Nov. 4 without cyber-attack training data
Google's EmbeddingGemma 2 Collapses Four Models Into 567MB of Phone RAM
The 740M open model maps text, code, images, video and audio into one vector space, and ships encoders you can leave on disk
Mistral's 1T 'Le Chonk' Sells a Refusal Gap, Not a Leaderboard Win
Mistral Large 4 is in preview with 49 billion active parameters, and the pitch is security work that closed models decline to do
Nvidia's Cheaper DGX Spark Has Half the Memory and Still Costs 25% More Than the Original
The 64GB model lands Oct. 23 at $4,999 — the same week the 128GB version went to $6,950, and the memory crunch is doing most of the work
A $300M Chip Smuggling Indictment Shows the DOJ Has Settled on a New Term: 'Super Intelligence'
Greg Lui is accused of routing export-controlled Nvidia servers through Malaysia and Singapore to Chinese buyers — and the Justice Department's announcement never once says 'AI'
A Stratego AI Went Superhuman on 16 GPUs After a $3 Million Attempt Fell Short
Ataraxos needed 34 times fewer self-play games than DeepMind's DeepNash and beat the game's most decorated player 15-1 — the gain came from an algorithm, not a budget
Ai2 Open-Sources an 8B Model That Cuts Cited-Report Time to 51 Seconds
AstaBrief-8B writes a whole scientific report in one pass instead of section by section — and Ai2 says it did not lose quality doing it
Apple Will Gate macOS Full Disk Access Behind 'Very Explicit User Action' — and Blames AI Agents
The permission that hands an app your mail, messages and browsing history in one click was built for backup tools. Apple says autonomous agents broke that assumption.
antirez's ds4 Runs DeepSeek V4 Flash at 34.4 Tokens/s on an M5 Max, 14.4 on a DGX Spark
The MIT-licensed C engine loads only its own GGUF builds — and the reference table inverts the hardware ranking depending on whether you count prefill or generation.
Pi 1.0 Ships the MCP Support Its Creator Once Called Unnecessary
Earendil folded the protocol into a sandbox called Codemode and spun long-running agents out into a separate package to keep Pi minimal