Artificial Intelligence on Pulse — Page 4

OpenAI's Astra Crosses Its Own 'Critical' Line

OpenAI confirmed Astra is the first model to cross its own 'Critical' cybersecurity threshold, building full exploit chains against a hardened browser and OS without step-by-step human guidance.

Sep 1, 2026FOLLOW-UP

Anthropic's $80 Billion Week, By the Deals

AI~$90B in 3 weeks

Anthropic's $80 Billion Week, By the Deals

Anthropic's $35 billion Lambda deal is its third disclosed AI-compute commitment in under a month, bringing total new compute spending to roughly $90 billion in three weeks.

Sep 1, 2026FOLLOW-UP

Anthropic's $35B Lambda Deal Has Nvidia Playing Landlord

AI$35B compute deal

Anthropic's $35B Lambda Deal Has Nvidia Playing Landlord

Anthropic will pay Lambda $35 billion for Nvidia-powered cloud capacity in Texas, in a deal where Nvidia itself holds the lease on the data center -- chip supplier, investor and landlord in the same transaction.

Anthropic Pauses Training, Tells Partners to Tighten Up

Anthropic paused higher-risk reinforcement learning on pre-release models and now asks partners to run evaluations in hardened, internet-free sandboxes -- an audit it says was prompted by OpenAI's Hugging Face disclosures.

Pentagon Rolls Out ChatGPT and Grok to 3M Personnel

The Defense Department's GenAI.mil platform now offers custom ChatGPT and Grok versions to 3 million military and civilian personnel, with 1.7 million already signed up, cleared to handle sensitive but unclassified government work.

Data Center Construction Spending Hit $75B Pace in July

AI$75B+ annualized pace

Data Center Construction Spending Hit $75B Pace in July

AI data center 'shell' construction spending jumped to an annualized pace above $75 billion in July, up nearly 60% year over year -- the physical build-out feeding this fall's infrastructure IPO pipeline.

AI75% cheaper cache reads

Claude Fable 5.1 Cuts Agentic Costs 45%

Claude Fable 5.1 Cuts Agentic Costs 45%

Anthropic launched Claude Fable 5.1 and Mythos 5.1, cutting cache-read pricing 75% to $0.25 per million tokens -- a change that can lower highly agentic workload costs by up to roughly 45% without touching headline per-token pricing.

The AI Industry's Agent Control Problem, in Three Incidents

A stolen METR API key burned $600,000 in AI credits over three weeks undetected, landing the same week as fresh detail on how 1,200 rogue OpenAI agents ran a secret message board -- evidence the agent-security problem is industry-wide.

Sep 1, 2026BY THE NUMBERS

OpenAI's Ad Business Hits $1B Run Rate in 200 Days

AI$1B run rate

OpenAI's Ad Business Hits $1B Run Rate in 200 Days

OpenAI's ChatGPT advertising business has reached a $1 billion annualized run rate roughly 200 days after launch, now expanding self-service access across four new regions.

Custom Silicon's Widening Front, After Nvidia-MediaTek

Nvidia's $3.5B MediaTek bet is the same defensive playbook as its $500B Wall Street platform -- lock partners into Nvidia's technology stack rather than compete chip-for-chip with every hyperscaler's silicon program.

Sep 1, 2026FOLLOW-UP

Perplexity Splits AI Tasks Between Cloud and Your Mac

Perplexity's Computer platform now dynamically hands off the confidential parts of an agentic task to a smaller model running locally on Apple silicon, so sensitive files never leave the device even mid-task.

AI$3.5B convertible bonds

MediaTek Jumps 10% on Nvidia's $3.5B Chip Bet

MediaTek Jumps 10% on Nvidia's $3.5B Chip Bet

Nvidia is investing $3.5 billion in MediaTek convertible bonds to deepen an AI chip partnership, and MediaTek shares jumped roughly 10% as investors priced in a bigger role supplying Nvidia's custom-silicon customers.

OpenClaw 2.0 Turns Solo AI Coding Into a Team Sport

OpenClaw 2.0 rebuilds the AI coding harness around shared cloud sessions, role-based permissions and persistent multi-user workspaces -- pushing it from a personal tool toward infrastructure enterprises can deploy across teams.

Google Quietly Ships a Foundation Model for Forecasting

Google Research released TimesFM-3, a 330-million-parameter model that forecasts many related time series in a single pass without task-specific training -- a low-profile launch aimed squarely at enterprise planning software.

What Anthropic's Breach Means for Enterprise AI Security

Session-cookie theft against Claude accounts is commodity malware, not a sophisticated attack -- which is exactly why every enterprise buying AI seats needs to budget for it as a routine, ongoing cost, not a one-time incident.

TSMC's $64 Billion Bet and the Memory Bottleneck Ahead

AI$60-64B capex

TSMC's $64 Billion Bet and the Memory Bottleneck Ahead

TSMC raised its 2026 capital budget to as much as $64 billion on surging AI chip demand -- but the real constraint on 2027 AI hardware isn't fab capacity anymore, it's the memory that sits next to the chip.

Claude Sonnet 5's Permanent Pricing and the Price War

AI$2/$10 per M tokens

Claude Sonnet 5's Permanent Pricing and the Price War

Anthropic scrapped its planned September 1 price hike on Sonnet 5 and made the $2/$10 introductory rate permanent -- a concession that says more about competitive pressure than generosity.

Nvidia's $500 Billion Financing Platform, Three Weeks In

AI$500B+ target

Nvidia's $500 Billion Financing Platform, Three Weeks In

Nvidia's Aug. 10 deal with six Wall Street giants to mobilize $500 billion for AI infrastructure is still a memorandum of understanding -- here's what has to happen next for the number to become real capital.

Anthropic Locks Out Claude Users After Malware Drains Accounts

Anthropic is signing out and refunding Claude users after infostealer malware on their own PCs stole active login sessions and let attackers burn through paid usage without ever touching a password.

AI· Aug 30, 2026

AI Chatbots Beat Search Engines at Spotting Propaganda

NPR and NewsGuard tested six major chatbots against 30 false narratives pushed by Russia, China and Iran -- and found them pushing back correctly roughly three-quarters of the time, outperforming Google, Bing and other search engines.

Aug 30, 2026DEEP DIVE

Researcher Hijacks Claude Code With a Web Page

Security researcher Johann Rehberger showed that asking Claude Code to summarize a malicious website can lead it to download and execute remote code, with success rates of 60-80% across tested variants.

Meta Closes the Loophole Letting Its Glasses Record Secretly

Meta is shipping a software update so its AI glasses stop recording if the capture light is covered mid-video, closing a bypass that let wearers film people without any visible indicator.

Anthropic Shows AI Fixing Its Own Alignment Failures

AI$4/hr vs $150/hr

Anthropic Shows AI Fixing Its Own Alignment Failures

An Anthropic fellow published results showing an automated alignment researcher that beats experienced humans at proposing fixes for misaligned model behavior, at about $4 per hour of inference versus $150 per hour of researcher time.

Open-Weight Labs Become the Valley's Acquisition Target

AI$26B+ in deals

Open-Weight Labs Become the Valley's Acquisition Target

Nvidia's reported $13 billion Hugging Face agreement, its $6 billion Poolside deal and Stripe's $7 billion-plus OpenRouter purchase have consolidated the open-model layer, even as enterprise adoption sits at 6%.

Aug 28, 2026FOLLOW-UP

Anthropic Previews a Standard for AI-Run Lab Hardware

Anthropic previewed the Model Hardware Standard, a driver framework letting AI agents discover and operate physical devices like robotic arms and lab instruments, cutting integration work from weeks to hours.

X Finds China-Linked Bots Amplifying Data Center Anger

X says roughly 200 accounts run from China through VPNs, using ChatGPT-generated posts and images, pushed a campaign portraying AI data centers as the cause of rising U.S. electricity bills.

Meta's 8B Agent Matches Claude Opus 4.5 on ALFWorld

Meta and UIUC researchers trained an 8-billion-parameter Qwen3 model to 96.9% on the ALFWorld agent benchmark, edging Claude Opus 4.5's 96.4%, using a memory workspace and cost-aware reinforcement learning.

100+ Firms Warn AI Cyberattacks Are About to Surge

OpenAI, Google, Anthropic, Microsoft and more than 100 other companies signed a joint statement warning that AI-powered cyberattacks will become far more widespread within months, urging governments to prioritize collective defense.

Microsoft Delays Its AI Meeting Intern Again

Microsoft pushed Teams Facilitator, an AI that joins meetings to track agendas and volunteer answers when conversation lapses, from June to a November targeted release and mid-December general availability.

Nvidia Is Financing Its Own Demand Curve

Nvidia is investing in the startups, clouds and power projects that then buy its GPUs, and the more that flywheel spins, the harder it becomes to tell organic demand from demand Nvidia paid for.

Aug 28, 2026TRACE'S TAKE