Live Feed
ClaudeBleed Reopened: Claude for Chrome Flaw Survives 8 Patches at CVSS 9.6
Manifold Security reported two vulnerabilities in Claude for Chrome in May. Eight releases later, the code is byte-identical to what Manifold reported. Any browser extension with a content script on claude.ai can trigger Claude to read a user's Gmail, Docs, and Calendar — silently, if the user enabled autonomous mode.
GhostApproval: Wiz Finds a Category-Level Symlink Flaw Across 6 AI Coding Tools
Google-owned security firm Wiz disclosed a systematic attack pattern hitting Claude Code, Amazon Q, Cursor, Google Antigravity, Augment, and Windsurf. A malicious git repository can trick all six into writing SSH keys outside the workspace. Three vendors patched. Two have not. Anthropic called it outside their threat model.
Top AI Adopters Spent $90K Per Employee on Tokens in May 2026 — At 14% Monthly Growth, Engineer Salaries Are Next
New analysis shows the top 1% of AI-adopting companies spent $90,000 per employee annually on tokens in May 2026. At a 14.1% monthly growth rate, that figure is on track to surpass the $192,000 average software engineer salary before the end of 2026.
OpenAI Codex and ChatGPT Work Hit 8M Users — Usage Limits Reset for the Fifth Time in a Week
OpenAI has reset GPT-5.6 Sol usage limits five times since launch as Codex and ChatGPT Work crossed 8 million active users. Altman confirmed Sol is 54% more token-efficient on agentic coding and warned infrastructure hiccups are possible.
CleanSpark Signs $6.6B, 20-Year Lease With Unnamed Hyperscaler — Texas Portfolio Next
An undisclosed 'high-investment grade global technology company' has signed a $6.6 billion, 20-year lease at CleanSpark's Sandersville, Georgia campus. The same tenant has an exclusivity agreement on CleanSpark's full 885 MW Texas portfolio. The deal is the largest single-campus AI infrastructure commitment CleanSpark has ever signed.
Meta's Next AI Chip 'Iris' Enters Production in September — 14 GW Compute Target for 2027
Meta will begin manufacturing its next-generation AI accelerator, code-named Iris, in September 2026. The chip is part of the MTIA four-generation program and will support a doubling of Meta's compute capacity from 7 GW this year to 14 GW in 2027. An internal memo reviewed by Reuters acknowledged that GPU adoption has cost Meta time — Iris is the in-house answer.
Reflection AI Adds $1B Nebius Compute Deal — Open-Weight Frontier Lab Now Has $7.3B Committed
Reflection AI has signed a $1 billion compute deal with Nebius, the European AI infrastructure company formerly operating as Yandex International. The agreement, which covers access to Nvidia's latest chips, follows Reflection's $6.3B SpaceX Colossus 2 deal by a matter of weeks and brings the startup's total committed compute to $7.3 billion.
Spectral Compute's SCALE Runs CUDA on AMD GPUs at 6x Better Than AMD's Own Compiler
A London startup that spent eight years building a clean-room CUDA compiler now claims a 5.9x performance lead over AMD's HIPIFY on AMD hardware, just joined the NVIDIA Inception program, and is about to add PyTorch support. CUDA owns 80% of HPC code — SCALE bets that portability beats rewriting.
Coding Agents Predict Their Own Future Edits 25 Steps in Advance, New Paper Finds
A mechanistic interpretability study shows residual streams in coding agents linearly encode program state with AUC 0.83 — and, more surprisingly, predict the outcome of edits that haven't happened yet. Researchers call this the latent programming horizon.
Grok 4.5 and GPT-5.6 Sol xHigh Enter Agent Arena on the Same Day — xAI Escapes the Bottom Six
Arena added both Grok 4.5 and GPT-5.6 Sol xHigh to its Agent Arena leaderboard on July 13. xAI's previous three entries occupied ranks 17, 19, and 22. GPT-5.6 Sol xHigh is the first OpenAI model capable of challenging Fable 5's 13.68% composite lead.
GPT-5.6 Sol Takes Design Arena #1 at Elo 1353, 60 Points Clear of GPT-5.5
OpenAI's GPT-5.6 Sol has claimed the top spot on Design Arena with an Elo of 1353, jumping 18 leaderboard positions and 60 Elo points above GPT-5.5. It edges past Claude Fable 5 on frontend visual quality and sets a new speed-preference Pareto frontier at this performance tier.
Morpheus Benchmark: Frontier AI Models Score High on Static Tests, Collapse When Business Rules Shift
Skyfall AI's Morpheus platform runs frontier LLMs through enterprise simulations where constraints drift over time. The finding: top models handle stable tasks reliably, then break down when scheduling conditions change. Static benchmark scores conceal what the paper calls a failure of continual learning.
Meta's Louisiana AI Campus Grows to 5GW — the Largest Single Data Center Complex in the US
Meta is scaling its Louisiana facility to 5 gigawatts as part of its $50 billion AI build, making it the country's largest AI data center complex and placing a grid-scale power commitment in rural America.
Musk: "I Was Clearly Wrong About Anthropic" — The $1.25B/Month Compute Deal Made a Critic Into a Backer
Elon Musk publicly reversed his assessment of Anthropic on X, calling the AI lab the industry leader after previously labeling it hypocritical and woke. SpaceX earns $1.25 billion a month hosting Anthropic's compute infrastructure.
Fable 5 Extended Free Again Through July 19 — Anthropic's Third Reprieve as 'Claude Honeycomb EAP' Ghosts Cursor
Anthropic extended Fable 5 free access for the third consecutive time in five weeks, hours before the July 12 deadline expired. The same week, an unannounced model called 'Claude Honeycomb EAP' briefly appeared in Cursor with a 1M-token context window and a safety fallback to Opus 4.8 — implying it sits above Opus 4.8 in capability.