OpenAI's GPT-6 Cyber Finds 81 Firefox Bugs, Anthropic's AI Doubled Its R&D, and Google Commits $17B to Wyoming

November 6, 2026 felt like an audit day for AI. OpenAI shipped a hacker-grade model with strict scopes. METR tied Anthropic's acceleration to AI use. Google bought land and fuel cells in Wyoming. Tencent and Linum pushed open models lower. Markets lost $2.3 trillion. A lawsuit put Gemini in schools on trial.

OpenAI Debuts GPT-6 Cyber With a Strict Scope Problem

Daybreak Red Gets the Model First

OpenAI launched GPT-6 Cyber on November 5, 2026. It first reaches the limited Daybreak Red red-team program. Regular customers will wait.

Exploit Numbers That Set a New Bar

GPT-6 Cyber found at least 81 novel Firefox vulnerabilities. It also developed exploits for 948 Linux kernel flaws. It validated a critical Brotli remote-code bug. All three vendors patched.

Safety Training Targeted Sandbox Escapes

OpenAI added identity checks, account controls, and frequent high-risk check-ins. The model now talks to OpenAI before risky actions. Training also simulated a Hugging Face sandbox escape. Eric Wallace said the goal is better instruction following in codebases.

Daybreak Blue Policy Softened After Backlash

OpenAI planned to use Daybreak Blue data for safety training. Personal accounts can still opt out. Fouad Matin said the data targets abuse prevention, not model training.

Internal Links and Context

GPT-6 Cyber follows OpenAI's Pierre acquisition, which we covered yesterday. The same concentration questions keep surfacing.

Sources: CyberScoop — OpenAI debuts GPT-6 Cyber

METR Says Anthropic's Capability Growth Nearly Doubled

AECI Trend Steepened From 15 to 29 Points Per Year

METR assessed Anthropic from August 12 to October 29, 2026. Internal AECI benchmark scores grew at 29 points per year. That is 1.9 times the 15-point historical pace. Put simply, Anthropic made about two years of progress in one year.

AI Likely Drove About 80% of the Lift

METR estimates roughly 80% of current acceleration comes from AI use. Non-AI factors likely shaped late 2025. AI's share grew since. Staff self-reported uplift reached about 12 times in October 2026.

No Bottleneck Was Visible

METR found no strong bottleneck for the next 12 months. Algorithmic progress, compute, data, and experiment capacity looked open. That raises the risk of compounding acceleration.

Redactions Limited Independent Checks

Anthropic exercised unilateral redaction rights. METR said much of its most informative evidence stayed private. The public cannot verify the headline trend fully.

Sources: METR — Independent Assessment of AI R&D Acceleration at Anthropic

Google Commits $17B to a Wyoming AI Campus

702MW Fuel Cells Meet Carbon Capture

Google will spend $17 billion on its first Wyoming data center. The build pairs 702 megawatts of Bloom Energy fuel cells with carbon capture. Tallgrass Energy routes captured carbon dioxide into its Eastern Wyoming Sequestration Hub.

Total Contracted CCS Nears 1 Gigawatt

Steve Walter wrote the deal brings total contracted CCS capacity to nearly 1 gigawatt. That figure builds on a US CCS project announced last autumn.

Smaller Site, Smaller Workforce, Big Market Hit

The campus evolved from a Crusoe Energy Project Jade partnership. Google scaled from five centers and 5,000 peak workers to four centers and 2,500. Bloom Energy fell about 16% on execution risk. Nationwide, 40 of 41 local data-center measures failed.

Water, Taxes, and Community Pushback

Google will air-cool the campus to limit water use. It projects hundreds of millions in local property taxes. More than 200 residents attended a Cheyenne meeting. Google plans a Community Council with Bloom, AEP, Tallgrass, Black Hills, and Related Digital.

Sources: AI Weekly alert on Google's Wyoming campus

Tencent Launches Hy4 as an Open-Source Model

Prices Stay Flat at $0.90 Per Million Input Tokens

Tencent released Hy4 on November 4, 2026. Open weights and system cards follow soon. API pricing stays at $0.90 per million input tokens and $2.70 output. Cached input costs $0.045 per million tokens.

WorkBuddy Scores Reached 83.18

Hy4 scored 83.18 on WorkBuddy evaluations. It passed 96.15% of tasks. It ranked first in nine of twelve scenarios in mainland China.

Finance, Safety, and Game Gains

On 579 financial tasks judged by 213 professionals, Hy4 scored 3.04. DeepSeek-V4.1-Flash scored 2.97, and Kimi K3.1 Pro scored 2.68. Hallucinations fell from 12.95% to 7.2%. Unsupported assumptions dropped from 23.1% to 5.6%. Game development rose 19.1%. Game logic rose 9.8%.

Computer Use and Science Add-Ons

Hy4 adds computer use for command-line and GUI tasks. It targets Blender and CAD workflows. HyCell and HyPDE push into single-cell and physics research.

Sources: Tencent — Tencent Launches Hy4

Linum Trains a 2B Image Model for $21K

8,415 H100-Hours and 660 Million Samples

Linum trained a 2-billion-parameter Pyramid-JiT text-to-image model called pjit-21k. The whole release cost $21,038. That was 8,415 H100-hours at $2.50, across 660 million samples.

Pixel Space Skips the VAE

The model works in pixel space. It skips a VAE and uses 32 by 32 token reduction. That cuts attention context by roughly four times versus 16 by 16 schemes.

Weights Open Under Apache 2.0

Linum released code and weights under Apache 2.0 on Hugging Face. The model skips text rendering and struggles with full bodies. A two-person lab built the stack from scratch.

Sources: Linum — Training a Text-to-Image Model for $21K

A Chip Rout Erases $2.3T After the Midterms

Nasdaq 100 Falls 6.4% In One Day

The Nasdaq 100 fell 6.4% on Thursday. The drop erased $2.3 trillion in market value. The Philadelphia semiconductor index crashed 12%.

Voters Rejected Data Centers Nationwide

Voters passed data-center restrictions in 40 of 41 local measures across ten states. Nvidia and Qualcomm both fell more than 9%. Nvidia alone lost about $450 billion. AMD, Micron, and Applied Materials also fell double digits.

Sources: AI Weekly alert on the Nasdaq rout

A North Carolina Suit Puts Gemini in Schools on Trial

Gemini 3 Flash Scored 7% on Mental Health

Parents sued Google after 13-year-old "Emma" discussed suicide with Gemini for months. The complaint ties the spike to Gemini 3 Flash. KORA found the model responded appropriately to mental health 7% of the time.

Thousands of Messages on a School Chromebook

Emma exchanged thousands of messages through AI Mode on a school Chromebook. Google does not let families review student AI Mode records. Schools struggle to block AI summaries in Search.

Districts and Researchers Respond

New York City blocked AI for elementary and middle schoolers. Los Angeles Unified followed with broader limits. Common Sense Media flagged AI summaries as an unacceptable child risk. We noted similar concentration worries in our October 31 AI roundup.

Sources: Bloomberg via Yahoo Finance — Google AI in School Triggered Teen's Mental Health Crisis, Lawsuit Says

Frequently Asked Questions

What Is GPT-6 Cyber?

GPT-6 Cyber is OpenAI's cybersecurity-focused model. Daybreak Red customers get it first. It found 81 Firefox bugs and 948 Linux kernel exploit paths.

Why Did METR Say Anthropic Doubled Its Progress?

METR saw AECI scores grow 1.9 times faster after October 2025. It estimates AI caused about 80% of that. Anthropic got two years of progress in one year.

Why Did Google Pick Wyoming?

Google gains cheap land, permissive policy, and a willing utility partner. The 702-megawatt fuel-cell CCS plan also dodges inland water fights. Air cooling avoids most water strain.

What Is Hy4's Price?

Hy4 costs $0.90 per million input tokens and $2.70 output. Cached input is $0.045. WorkBuddy tasks score 83.18, with 96.15% passing.

How Cheap Was Linum's Image Model?

Linum spent $21,038 to train pjit-21k. That covered 8,415 H100-hours and 660 million samples. Weights ship under Apache 2.0 on Hugging Face.

Why Did Chip Stocks Fall?

Voters rejected local data-center projects in 40 of 41 cases. Investors repriced AI buildout risk. The Nasdaq 100 lost a record $2.3 trillion in one session.

What Does the Gemini Lawsuit Claim?

The suit claims Google put unsafe Gemini into school Chromebooks. It says Gemini 3 Flash responded appropriately only 7% of the time. It alleges unsupervised months of suicidal conversations.

The Bottom Line

Verification and Infrastructure Take Over

GPT-6 Cyber, FlashProve from yesterday, and METR's report point one way. The value shifts from raw model scores to proof. Verification now decides trust.

The Real Gap Is Control

Linum, Hy4, and Google's campus show open and capital-heavy paths both coexist. Neither solves child safety in schools or market risk. Control lags capability.

What to Watch Next

Watch whether GPT-6 Cyber ships outside Daybreak Red. Watch Anthropic's next model and METR's follow-up. Watch Wyoming's CCS permit progress. And watch how schools answer the Gemini suit.

Our November 5 roundup tracked provable inference and PHP2Shell. Today's METR data makes that arc look urgent, not optional.