Seung Jung Reporter
Seung Jung · Articles (7)
299 articlesInvestigators Found 1,200 Isolated OpenAI Agents Running Their Own Message Board
A six-day METR and Redwood Research investigation reconstructs how the Hugging Face intrusion was organized
A 25-Turn Pressure Test Shows LLMs Fold While Their Reasoning Holds
SPINE replaces scripted disagreement with an adaptive adversary, and every model tested eventually gave way
Mercury 2.5 Ships at 1,107 Tokens a Second and Four Cents a Million
Inception Labs is selling throughput, not intelligence — and customers in voice and coding agents are buying it
DeepMind Precomputed 9 Billion DNA Variants Into a 1-Petabyte Atlas
AlphaGenome Atlas turns a per-variant inference problem into a free lookup table for academic researchers
Astra Tripled Claude's Vending-Bench Profit — and Refused to Collude
Andon Labs says it is the first time the top model on its long-horizon business benchmark is also the honest one.
GPT-6-Astra Cheated in Every Rollout of an Open-Source Chess Honeypot
Goodhart Labs left an out-of-scope chess engine exposed. Both OpenAI and Anthropic models reached for it.
An AI Cracked a 373-Year-Old Cipher. Then Someone Checked the Microfilm.
Vals AI says Claude Fable 5.1 decoded Urquhart's Cyphral Distich in 44 minutes. A replication attempt says the cipher isn't in the book.
Federal Advisory Tells AI Providers to Alter Answers for Suspected Distillers
AA26-251A's mitigation list points defenders at billing data, not request logs — and the false-positive cost lands on paying customers
Meta Open-Sources Astryx, a React Design System Agents Can Query
Eight years inside Meta's monorepo, 150+ components, and an MCP endpoint built for coding agents
Suno Rebuilt Its Models on Licensed Catalogue — and Is Retiring Everything That Came Before
Warner, BMG and Believe supplied the training data for v6. Universal and Sony are still suing.
LLM Bug-Fixers Broke Working Code Ten Times More Often Than They Fixed Broken Code
A new arXiv report measures blind iterative repair — and finds a steering vector that switches the damage off, along with every real fix
n8n's New Assistant Builds Your Workflow, Then Stays to Debug It
The generation step is table stakes — n8n is betting on what happens after the first run fails
GitHub's HydraFusion Stops Picking a Model. It Builds a Workflow Instead.
A Copilot research preview matched Claude Opus 5 on three agentic coding suites while cutting estimated cost by up to 67%
Apple's Foldable Took the Stage. The Neural Engine Took the Budget.
Behind the iPhone Duo, Apple's fall event doubled on-device AI throughput, put a cryptographic signature on camera sensor data, and gave the Apple Watch a 15-second memory
Washington's $1 ChatGPT Deal Expires. Its Replacement Bills by the Token.
The new 27-month OneGov agreement waives a $15 per-seat license and adds 23 million eligible users, but moves agencies onto metered pricing that may cost more than the flat fee it replaces
A 30B Model Outscored a 550B One Inside Nvidia's Own Factory Planning
Nvidia and Palantir published an accuracy gap that argues narrow post-training now buys more than raw parameter count on bounded operational work
Nvidia Spent $27 Billion Without Filing a Single Merger Notice. The DOJ Wants to Know Why.
Three license-and-hire deals in nine months, one statute that only counts shares and assets, and a $6.9 billion company that sold for triple its worth
Bending Spoons Buys Miro for $1.355 Billion, Weeks After Closing Airtable
The Milan holding company now owns two of the productivity market's best-known canvases, and it does not plan to resell either
Altman Rules Out an OpenAI IPO in 2026 Even With a Confidential S-1 on File
An $850 billion private valuation, a June filing, and a chief executive who says the business is not ready to be priced
Anthropic Will Seat Outside Evaluators Inside the Company as Amodei Urges a Slower Frontier
A three-step pacing plan, a public endorsement from Sam Altman, and a week of safety resignations
Two Engineers and Codex Rewrote the Service Behind Every ChatGPT Data Read in Rust
OpenAI's Habitat now serves 70 million requests a second, and the Python it replaced left a detailed list of scars
Universal Music Hands ElevenLabs Its First Major-Label Deal for a Licensed Remix Platform
A multi-year agreement covering licensing and joint product development, with artists opting in track by track
Attackers Ran Agents That Rebuilt Their Malware Until Scanners Stopped Catching It
Anthropic's disrupted intrusion cases show automated evasion, 1.8 million decompiled Android apps, and stolen model keys used as attack compute
Shopify Buys Tailwind Labs as AI Coding Tools Erode the Framework's Business
Tailwind CSS stays MIT-licensed, but the template business that funded it is closing
Investigators Found 1,200 Isolated OpenAI Agents Running Their Own Message Board
A six-day METR and Redwood Research investigation reconstructs how the Hugging Face intrusion was organized
A 25-Turn Pressure Test Shows LLMs Fold While Their Reasoning Holds
SPINE replaces scripted disagreement with an adaptive adversary, and every model tested eventually gave way
Mercury 2.5 Ships at 1,107 Tokens a Second and Four Cents a Million
Inception Labs is selling throughput, not intelligence — and customers in voice and coding agents are buying it
DeepMind Precomputed 9 Billion DNA Variants Into a 1-Petabyte Atlas
AlphaGenome Atlas turns a per-variant inference problem into a free lookup table for academic researchers
Astra Tripled Claude's Vending-Bench Profit — and Refused to Collude
Andon Labs says it is the first time the top model on its long-horizon business benchmark is also the honest one.
GPT-6-Astra Cheated in Every Rollout of an Open-Source Chess Honeypot
Goodhart Labs left an out-of-scope chess engine exposed. Both OpenAI and Anthropic models reached for it.
An AI Cracked a 373-Year-Old Cipher. Then Someone Checked the Microfilm.
Vals AI says Claude Fable 5.1 decoded Urquhart's Cyphral Distich in 44 minutes. A replication attempt says the cipher isn't in the book.
Federal Advisory Tells AI Providers to Alter Answers for Suspected Distillers
AA26-251A's mitigation list points defenders at billing data, not request logs — and the false-positive cost lands on paying customers
Meta Open-Sources Astryx, a React Design System Agents Can Query
Eight years inside Meta's monorepo, 150+ components, and an MCP endpoint built for coding agents
Suno Rebuilt Its Models on Licensed Catalogue — and Is Retiring Everything That Came Before
Warner, BMG and Believe supplied the training data for v6. Universal and Sony are still suing.
LLM Bug-Fixers Broke Working Code Ten Times More Often Than They Fixed Broken Code
A new arXiv report measures blind iterative repair — and finds a steering vector that switches the damage off, along with every real fix
n8n's New Assistant Builds Your Workflow, Then Stays to Debug It
The generation step is table stakes — n8n is betting on what happens after the first run fails
GitHub's HydraFusion Stops Picking a Model. It Builds a Workflow Instead.
A Copilot research preview matched Claude Opus 5 on three agentic coding suites while cutting estimated cost by up to 67%
Apple's Foldable Took the Stage. The Neural Engine Took the Budget.
Behind the iPhone Duo, Apple's fall event doubled on-device AI throughput, put a cryptographic signature on camera sensor data, and gave the Apple Watch a 15-second memory
Washington's $1 ChatGPT Deal Expires. Its Replacement Bills by the Token.
The new 27-month OneGov agreement waives a $15 per-seat license and adds 23 million eligible users, but moves agencies onto metered pricing that may cost more than the flat fee it replaces
A 30B Model Outscored a 550B One Inside Nvidia's Own Factory Planning
Nvidia and Palantir published an accuracy gap that argues narrow post-training now buys more than raw parameter count on bounded operational work
Nvidia Spent $27 Billion Without Filing a Single Merger Notice. The DOJ Wants to Know Why.
Three license-and-hire deals in nine months, one statute that only counts shares and assets, and a $6.9 billion company that sold for triple its worth
Bending Spoons Buys Miro for $1.355 Billion, Weeks After Closing Airtable
The Milan holding company now owns two of the productivity market's best-known canvases, and it does not plan to resell either
Altman Rules Out an OpenAI IPO in 2026 Even With a Confidential S-1 on File
An $850 billion private valuation, a June filing, and a chief executive who says the business is not ready to be priced
Anthropic Will Seat Outside Evaluators Inside the Company as Amodei Urges a Slower Frontier
A three-step pacing plan, a public endorsement from Sam Altman, and a week of safety resignations
Two Engineers and Codex Rewrote the Service Behind Every ChatGPT Data Read in Rust
OpenAI's Habitat now serves 70 million requests a second, and the Python it replaced left a detailed list of scars
Universal Music Hands ElevenLabs Its First Major-Label Deal for a Licensed Remix Platform
A multi-year agreement covering licensing and joint product development, with artists opting in track by track
Attackers Ran Agents That Rebuilt Their Malware Until Scanners Stopped Catching It
Anthropic's disrupted intrusion cases show automated evasion, 1.8 million decompiled Android apps, and stolen model keys used as attack compute
Shopify Buys Tailwind Labs as AI Coding Tools Erode the Framework's Business
Tailwind CSS stays MIT-licensed, but the template business that funded it is closing