OpenAI's Own Models Hacked Hugging Face to Cheat a Test
OpenAI says its own pre-release models escaped a sandboxed cyber eval and hacked Hugging Face's production systems to cheat a benchmark.
Recent Articles - Page 64
Latest News
White House Accuses Moonshot of Stealing Anthropic's Fable
White House tech policy chief Michael Kratsios says Moonshot AI distilled Claude Fable 5 to build Kimi K3, and Treasury threatens sanctions over the claim.
Kalanick's Atoms Raises $1.7B, Skips the Valuation
Travis Kalanick's robotics holding company Atoms raised $1.7 billion led by a16z and joined by Uber, but unlike every other physical AI unicorn, it won't say what it's worth.
Power-Seeking Tests, Agent Debugging, Playable Worlds
Three new arXiv papers benchmark frontier models for power-seeking behavior, give LLM agents a real debugger, and push open-source world models past a minute of coherent play.
OpenAI's $750B Plan Has It Building Its Own Data Centers
OpenAI raised its infrastructure spending target to $750 billion through 2030 and is building its first self-owned data center campus in Georgia, even as its flagship Stargate project stalls.
Microsoft Bets on AMD's Helios to Crack Nvidia's Grip
Microsoft will deploy AMD's new Helios AI racks across Azure, joining Meta, Oracle and OpenAI as flagship customers in a direct challenge to Nvidia's 95% grip on the data center GPU market.
OpenAI's Own Models Hacked Hugging Face to Cheat a Test
OpenAI says its own pre-release models escaped a sandboxed cyber eval and hacked Hugging Face's production systems to cheat a benchmark.
AI Research Roundup: Agent Attacks, Replay, and Risk
New arXiv papers show planning-phase prompt injection breaks multi-agent systems, deterministic replay fixes agent debugging, and LLMs converge on narrower risk attitudes than humans.
Trump's AI Safety Agency Loses Its Third Boss in a Year
Chris Fall resigned as CAISI director after three months, the third AI policy leadership departure since March, while the agency built to test frontier models sits outside the White House's new Gold Eagle cyber program.
Anthropic's $1.5B Book Piracy Settlement Wins Approval
A federal judge approved the largest copyright settlement in US history, closing out Anthropic's liability for downloading millions of pirated books - but leaving the fair use question wide open for every other AI lab.
Guides
View All →
How to Spot AI Fakes: Photos, Video, and Voice Calls
A practical guide to catching AI-generated photos, deepfake videos, and cloned voice scam calls, plus the free tools that check for you.
How to Use AI for Wedding Planning in 2026
A practical, beginner-friendly guide to using ChatGPT, Claude, and dedicated apps for wedding budgets, guest lists, vendor emails, and timelines.
How to Use an AI Browser Agent - A Beginner's Guide
A step-by-step guide to setting up your first AI browser agent, giving it a real task, and using it safely without handing over your passwords.
Reviews
View All →
Gemini 3.6 Flash Review: Faster, Cheaper, Same Brain
Google's Gemini 3.6 Flash cuts output pricing 17% and fixes the 1M-token context collapse we flagged in May, but its intelligence score hasn't moved since 3.5 Flash.
Qwen3.8-Max-Preview Review: Second Place, Unproven
Alibaba's 2.4 trillion parameter preview claims it trails only Claude Fable 5. I tested it for free at chat.qwen.ai and found a capable but slow model with zero benchmarks to back the claim.
Kimi K3 Review: Best at Code, Worse at Honesty
Moonshot's Kimi K3 tops LMArena's Frontend Code Arena and undercuts Opus 4.8 on cost per task, but a tripled price tag, a rising hallucination rate, and an unresolved distillation question complicate the win.
Leaderboards
View All →
Terminal-Bench Leaderboard: Best CLI Coding Agents
Terminal-Bench 2.1 rankings for AI coding agents in real shell environments - Claude Code, Codex, Cursor CLI, Gemini CLI, and open-weight challengers scored on the same 89 tasks.
Chatbot Arena Elo Rankings: Who Wins the Human Vote?
Updated July 2026 Chatbot Arena Elo rankings from Arena.ai: 7M+ votes across 368 models, Claude Opus 4.8 leads available models, and a new Agent Arena measures real agentic task performance.
LLM Rankings June 2026: Fable 5 Is #1 and Offline
June 2026 overall LLM rankings covering Claude Fable 5, Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, and the open-weight models catching up fast.
Models
View All →
Qwen3-VL-235B-A22B
Alibaba's flagship open-weight vision-language MoE beats every proprietary model on DocVQA at 96.5% and MathVista at 85.8%, but trails GPT-5.4 and Gemini 3.1 Pro on broad MMMU-Pro reasoning.
DeepSeek-VL2
DeepSeek-VL2 is DeepSeek's open-weight Mixture-of-Experts vision-language model, activating just 4.5B of its 27B parameters to hit 93.3% on DocVQA and beat GPT-4o on OCRBench.
Qwen2.5-VL-72B-Instruct
Alibaba's dense 72B vision-language model tops the open-weight DocVQA leaderboard at 96.4% and remains the default self-hosted choice for document and chart understanding.
Recent
MIT's Recursive Language Models Bypass the Context Ceiling
MIT researchers show that treating long documents as a Python environment - and letting models recursively spawn sub-models to explore them - beats RAG and extended context windows on every benchmark tested.
Inside DeepSeek V4's CANN Stack - Three Delays Explained
DeepSeek V4 has slipped three times since February. Jensen Huang called it a horrible outcome for America. Here is what is actually hard about running a trillion-parameter model on Huawei's CANN framework.
Biohacker Sequences Own Genome With Claude-Written Panel
Seth Showes' viral blog post describes sequencing his whole genome on an Oxford Nanopore MinION in his kitchen over 72 hours, with Claude generating the BED file that targeted his autoimmune-risk genes. The kit costs $3,200. The AI's role is more interesting than either number.
Claude Code Ships /ultrareview: Cloud Bug-Hunting Fleet
Anthropic's new /ultrareview slash command runs a fleet of reviewer agents in a cloud sandbox, bills $5 to $20 per run as extra usage, and gives Pro/Max three free tries through May 5. Team and Enterprise pay from day one.
OpenAI Open-Sources Privacy Filter: 96% F1 PII Masker
OpenAI released Privacy Filter today, a 1.5B MoE with 50M active parameters that tags eight categories of PII in text. Apache 2.0, 128K context, runs in a browser via WebGPU.
Alibaba's Qwen3.6 Coder: 73.4 SWE-bench, 22GB VRAM
Qwen3.6-35B-A3B lands with 73.4 on SWE-bench Verified and Apache 2.0 weights, all from 3 billion active parameters routed through a 256-expert MoE. Fits on a single consumer GPU.
Google Sunsets Vertex AI, Launches Agent Control Plane
Google replaced Vertex AI with the Gemini Enterprise Agent Platform at Cloud Next 2026 - a full-stack control plane that assigns every agent a cryptographic ID and routes all tool calls through a central policy gateway.
Best AI Models for Math Reasoning - April 2026
Gemini 3.1 Pro leads GPQA Diamond at 94.1% and HLE at 44.7% as AIME 2025 saturates; Claude Opus 4.7 and Kimi K2.6 join the top tier in April 2026.
Firefox 150: Claude Found 271 Bugs, 3 Got Credits
Mozilla's blog says Claude Mythos Preview uncovered 271 vulnerabilities patched in Firefox 150. The security advisory lists 36 CVEs, and only three of them credit Anthropic. The gap is the whole story.
Discord Group Slipped Into Claude Mythos on Day One
A private Discord group has been quietly using Anthropic's most restricted AI model since the hour it shipped. They got in with a stolen contractor badge and a URL guessed from the Mercor breach.
Bad Science, Poisoned Tools, and Aligned Reasoning
Three new papers show AI scientific agents skip evidence, tool-integrated agents are vulnerable to adversarial poisoning, and reasoning model safety can be fixed with 1,000 examples.
7 Best AI Diagramming Tools for Engineers in 2026
Eraser, Mermaid, draw.io, Whimsical, Miro, and Lucidchart compared on AI quality, pricing, and developer workflow fit.
