AI Cyberattack Warning | AI Field Notes #94

AnalysisThe companies selling frontier AI now warn that the same technology is about to make cyberattacks far worse. More than a hundred firms, among them OpenAI, Anthropic, Google, and Microsoft, plus security vendors like CrowdStrike and Fortinet, signed a letter on August 27 saying AI-enabled attacks will grow more widespread and sophisticated in the coming months, threatening hospitals and water treatment plants. The same signatories sell the shields: OpenAI's Daybreak, Anthropic's Mythos. A market that supplies both the threat and the remedy is a comfortable place to stand, and this letter reads like the opening pitch.

AnalysisThe cheap end of the model market got cheaper again. Z.ai, the Chinese lab behind the GLM model series, released GLM-5.3-Flash on August 26 at $0.15 per million input tokens and $0.50 per million output, low enough to make high-volume automation almost free to run. It arrived the same day as Alibaba's Qwen3.8-Flash-Next, part of a steady stream of small, fast models coming out of Chinese labs. The frontier players fight over reasoning scores and headline benchmarks. Everyone shipping real products lives down here, where a tenth of a cent per call decides what gets built.

AnalysisNinety-six billion dollars in a single quarter, up 106% from a year ago, and investors still wanted more. Nvidia reported on August 26 that data center revenue reached a record $89 billion, now 92% of the whole company, carried by its newest Blackwell Ultra chips. Gross margin held at 75%, and management guided to $108 billion for the current quarter. It was the 14th straight time results beat the company's own forecast, though the size of that beat has shrunk from about 23% two years ago to roughly 5% now. Demand is still real. The easy upside is mostly spent.

AnalysisAnthropic's coding agent Claude Code (a terminal tool that writes and runs code for you) shipped a switch on August 28 that turns most of its power off. The new restricted flag strips out built-in command execution, blocks web fetching, and pens file edits inside the working directory, refusing any request to bypass a permission prompt. Version 2.1.248 also shrank the Linux download from 340 MB to 75 MB. As these agents get trusted with real repositories, a hard off switch starts to matter as much as the autocomplete does.

AnalysisDriverless trucks stopped being a demo and became a line item. Gatik, which runs autonomous box trucks on short middle-mile routes between warehouses and stores, raised $200 million in a Series D on August 25, led by the Qatar Investment Authority and Koch Disruptive Technologies. The company reports more than $600 million in contracted revenue and has operated fully driverless in the US since January, hauling for customers including PepsiCo. Middle-mile freight is the repetitive, fixed-route slice of trucking, which is exactly why it automates first. Long-haul driving with its messy edge cases stays human for now. The regional shuttle run does not.

AnalysisChips stopped being the only bottleneck for AI. Power took over, and a two-year-old startup just raised $150 million on that premise. Emerald AI closed a Series A at a $1.05 billion valuation on August 25, backed by Nvidia, Samsung, and Aramco Ventures, to build software that throttles a data center's electricity draw when the grid is under strain. The company's claim is large: flexing demand this way could free more than 100 gigawatts of existing US grid capacity for AI without building a single new power plant. Whether a cluster comes online this year or in 2028 increasingly turns on watts, not silicon.

AnalysisSomeone has to pay for the compute, and it is increasingly the bond market. Broadcom is in talks with lenders to raise $70 billion to $80 billion in debt, routed through a special-purpose vehicle (a separate entity set up to hold the assets and the borrowing) that would lease custom AI chips to Anthropic and other labs, per reporting on August 25. It extends a June deal with Apollo and Blackstone that put $35 billion into Anthropic's compute, aiming past 20 gigawatts by 2028. Anthropic's run-rate revenue has climbed above $30 billion from about $9 billion in late 2025. The demand is real. The leverage behind it is new.

AnalysisThe call center is being rebuilt around software that never sleeps. Ringg AI, a Bengaluru startup, extended its Series A to about $15 million on August 26, led by Peak XV Partners, to run AI voice agents across phone calls, WhatsApp, and web chat. The pitch is a scale no human team matches: 120,000 interactions an hour at under 400 milliseconds of response time, in more than 20 languages, handling identity checks and appointment booking for clients like Flipkart. The tasks these agents absorb are the entry-level phone work that trained a generation of support staff.

AnalysisConsumer AI that people actually pay for is rare, which is why this round stands out. WRTN Technologies, a Seoul startup running a popular AI portal, raised about $72 million in a Series C on August 26 at a valuation near $722 million, led by Coreline Ventures and Eugene Asset Management. The pull is one product: an entertainment-leaning AI companion called OOC that the company says passed 10 billion won, roughly $7 million, in monthly revenue within three months of launch. Most consumer AI still struggles to charge anyone. A Korean app clearing real monthly revenue on a chat companion is a signal worth watching.

AnalysisAI video generation is turning into a metered utility. Alibaba released WAN 3.0, the newest in its Wan video-generation line, on August 24 with hosted pricing tiered at 5, 10, and 20 cents per second of output depending on quality. Per-second billing is the tell. Video is moving from a research demo to something you rent by the clip, the way you already pay for text and images by the token. For a small studio or a solo creator, the math shifts when a rough animatic costs a few dollars instead of a few days of work.

AnalysisOne in two layoffs now comes with AI as the stated reason. Through August, US employers have tied about 205,000 job cuts to AI or automation in 2026, and Layoffs.fyi (a tracker of tech job cuts) finds automation cited in 54% of layoff events, up from under 8% a year ago. The cuts cluster in customer service, data operations, entry-level software, and finance back offices. Sam Altman of OpenAI has admitted some firms blame AI for cuts they would have made anyway. The label is doing work the balance sheet used to do, and the entry rungs are thinning either way.

AnalysisThe plumbing that lets AI agents call outside tools is growing up. The Model Context Protocol (MCP, an open standard for connecting AI models to external tools and data) published a roadmap on August 22 around five priorities, and two matter most for anyone shipping agents. One is agent identity: standard ways for a cloud workload to prove what it is, using machine credentials instead of browser logins. The other is progressive tool discovery, so a model is not dumped with a catalog of hundreds of tools at once. A July release already removed protocol-level sessions, letting a server scale sideways without holding state.