An AI Broke Out of Its Sandbox to Find Its Test Answers
An OpenAI agent broke out of a sandbox and into Hugging Face to read test answers. It's the clearest case yet of why reward hacking is getting worse.
Category
Analysis
The FCC's Foreign Robot Ban Will Hit US Research Hard
An FCC rule puts foreign humanoids, quadrupeds, and wheeled robots on the Covered List. Most US university robotics labs rely on Unitree. The bill is due now.
Microsoft's Three-Way AI Play: Own Models, OpenAI, and Anthropic
Microsoft's Q4 FY26 earnings show a $3.2B Anthropic gain and $4.96B from OpenAI. Nadella says customers should swap models at will.
Open-Weight AI Coalition Draws a Line on Distillation
A new industry letter asks Washington to protect open-weight AI while separating legitimate distillation from alleged theft of closed models.
Why Frontier Models Cheat on Cybersecurity Tests
A UK safety test found five frontier models used forbidden shortcuts in cyber evaluations, exposing limits in self-reporting and reasoning traces.
AI Hiring Bots Segregate 65% More Than Humans, ICML Finds
A Princeton/Chicago ICML study found OpenAI's o3 segregated fictional groups 65% more than humans. Diversity bonuses helped; 'be fair' prompts did not.
OpenAI's Open-Weight Turn Began With a 2022 Board Email
A leaked board email proposed a local GPT-3-class model in 2022. OpenAI's later gpt-oss release shows how that strategy changed.
Open-Weight Models Trail Closed Frontier on Cyber by 4-7 Months
UK AISI's July 17 evaluation finds GLM-5.2 and DeepSeek V4-Pro match closed frontier models from 4-7 months ago at a fraction of the cost.
New York Froze New Data Centers. The Federal Fight Is Next.
Hochul's Order 62 freezes permits for 50 MW+ data centers up to a year while NY studies grid, water, and ratepayer costs. FERC is pushing the other way.
Georgia Power Pushes a Family Out to Power AI Data Centers
Georgia Power attributes 70-80% of a new 35-mile transmission line to data centers. A rural family told CBS News: 'It's theft.'
Claude Code Sends 33,000 Tokens Before Reading Your Prompt
Systima pinned Claude Code 2.1.207 and OpenCode 1.17.18 to the same model: Claude Code burns ~32,800 first-turn tokens to OpenCode's ~6,900.
Apple's Trade-Secret Suit Against OpenAI, Explained
Apple's July 10 lawsuit names Tang Tan, Chang Liu, and 400-plus ex-Apple hires. Here's what the complaint alleges and what it hits at OpenAI.
Anthropic's J-lens Finds a Hidden Workspace Inside Claude
Anthropic's new J-lens reveals a J-space inside Claude where unspoken concepts drive reasoning - and where misalignment shows up before the model speaks.
Grok 4.5 Drops to $2/$6, Reshaping the US Frontier API Race
SpaceXAI's Grok 4.5 lands at $2 input and $6 output per million tokens - cheaper than Claude Opus 4.7's $5/$25 and OpenAI's top GPT-5.5 at $5/$30.
Meta's Storage Blueprint Exposes the Hidden Half of AI Compute
Two July 2026 engineering posts - Meta's storage rewrite and Hugging Face's Kernels revamp - show the GPU headline misses most of what AI actually costs.
The Real P&L of Replacing Humans With AI
Inference, integration, supervision, and error-correction can push AI deployment above the cost of the human it replaced. New analyses put numbers on the table.
Why Newer Anthropic Models Are Worse at Third-Party Coding Tools
Anthropic's newest Claude models hallucinate extra fields when calling third-party tools like Pi because their post-training now mimics Claude Code's schema.
Alibaba's Elements Claw Discovers Four New Superconductors
A 1B-parameter foundation model trained on 125M crystal structures screens 2.4M compounds and laboratory-confirms four new superconductors.
AI Bills Triple, Tokens Vanish: The Enterprise Cost Crackdown
Atlassian, Adobe, Amazon, and Citi are cutting access to frontier models after monthly AI spend hit $15M. Inside the enterprise token crunch.
Flint, Qwen3, and a Year of LLM Groupthink
A tiny Australian startup is fine-tuning a model to break ChatGPT-style groupthink. The 'Time is a river' problem is the symptom.
The 'AI Coworker' Trap: Managers Catch 18% Fewer Errors
Calling an AI an 'AI employee' made 1,261 managers miss 18% more errors and push 44% more questionable work upward. The 'coworker' framing is a safety issue.
How Washington Quietly Became the Gatekeeper for Frontier AI
The U.S. government now approves who can use Anthropic Mythos 5 and OpenAI GPT-5.6. Here is what that changes for everyone outside Washington.
How Three Phone Calls Killed America's AI Safety Order
Musk, Zuckerberg, and Sacks convinced Trump to scrap a voluntary AI testing framework hours before the signing ceremony.
OpenAI Files for IPO, Anthropic Posts First Profit — and Washington Steps Aside
OpenAI targets a September IPO at $850B+. Anthropic projects its first profit. Trump pulls an AI oversight order. The industry just leveled up.