Tech Digest – July 23, 2026
AI Espionage & Trade
First AI Espionage Scandal — White House Accuses Moonshot AI of Distilling Anthropic’s Fable, Treasury Threatens Sanctions
The White House revealed that Moonshot AI built a covert internal platform to systematically distill Anthropic’s Fable model into its Kimi K3, using GB300 chips tapped in Thailand to evade detection. It is the first time a senior US official has publicly named a specific Chinese lab for copying a specific American model. Treasury Secretary Scott Bessent escalated immediately: “Open source is not open season on American IP,” with sanctions and Entity List designations on the table — the same punishment applied to Huawei in 2019. Anthropic had previously traced approximately 3.4 million Claude exchanges to Moonshot AI.
The reaction split fast. Jensen Huang called the Chinese models “excellent,” argued free AI sells hardware, and urged Anthropic to unshackle its Mythos model family. Nearly 200 startups, organized as the Little Tech Association, warned Washington that restricting access to Kimi and Qwen would kill hundreds of companies built on open-weight Chinese models. Even OpenAI’s Greg Brockman conceded K3 is “a pretty good model.” Commentators noted the irony: Anthropic trained on the entire corpus of human knowledge and sells it back compressed — and now objects to compression.
Note: Yesterday in this digest, Kimi K3 was benchmarked as competitive with Fable at a fraction of the cost. Today, the White House says it knows why. The sanctions threat creates immediate procurement risk for anyone building on Chinese open-weight models — but the 200-startup coalition reveals how deep the dependency runs. For EU institutions evaluating open-weight alternatives to reduce closed-model lock-in, the calculus just acquired a third variable: geopolitical supply-chain risk.
Sources: White House (Kratsios), Treasury (Bessent), Axios, Politico, TechCrunch
Cybersecurity Under Pressure
432 LLM-Found CVEs in One Weekend Bury Linux Kernel Maintainers
The Linux kernel project published 432 CVEs between July 19 and 20 — on top of more than 40 already issued that month. Kernel maintainer Greg Kroah-Hartman warned the flood is not over: “The number of LLM-found issues is only on the rise right now. It’s going to be a very long 18 months at the least to dig ourselves out of this mess.” Organizations that once reviewed each kernel CVE and backported selectively now face a volume that collapses the model entirely.
Note: Every cloud instance, every server, every container in institutional infrastructure runs a Linux kernel. When automated vulnerability discovery outpaces human review capacity, the patch management model that enterprises have relied on for two decades breaks. The maintainer gave you the timeline: 18 months, at least.
Sources: oss-sec (Greg Kroah-Hartman), oss-sec (CVE list)
AI at Operational Scale
Gemini Reaches 950 Million Monthly Users as OpenAI Ships Enterprise Voice Agents
Google’s Gemini now serves 950 million monthly active users — approaching a billion people interacting with AI tools every month. OpenAI, meanwhile, launched Presence, an enterprise platform for deploying AI voice and chat agents with built-in guardrails, policy enforcement, and evaluation tools. Presence already handles OpenAI’s own English-language support line, resolving 75% of inbound calls without human intervention. Within weeks of deployment, it met or exceeded benchmarks used to grade frontline human support quality, and a Codex-powered feedback loop reduced human handoffs by an additional 15 percentage points in 10 days.
Separately, Robinhood customers have begun delegating research and trade execution to AI agents — a shift from using AI as an information tool to using it as a decision-maker.
Note: When the company selling the AI tool uses it to automate 75% of its own support calls within weeks, that is not a marketing demo — it is a deployment benchmark. For any institution running a citizen inquiry line or service desk, the question is no longer whether AI agents can handle the volume.
Sources: The Verge, OpenAI, The Information
Research Automation
AI Cracks Decades-Old Mathematics as the US Commits $5 Billion to the Genesis Mission
Cognition’s Devin solved three graph theory conjectures open for 20 to 40 years — refuting Graffiti Conjecture 154 and proving Graffiti Conjectures 39 and 40 — in a single day, triggered by a tweet. Separately, a researcher working with GPT-5.6 Pro reported refuting the 30-year-old Dinitz-Garg-Goemans conjecture on graph flows, a problem that had occupied many of the field’s leading experts for decades.
The US government is treating this as a paradigm, not a series of anecdotes. The Genesis Mission, launched by executive order in November 2025, expanded this week into a $5 billion, 15-agency national program. It selected 278 projects from over 5,000 applications spanning all 50 states, backed by a consortium of 157 companies, 142 universities, and 16 national laboratories. The stated goal: double the productivity and impact of American research within a decade.
Note: Yesterday’s digest: an 87-year-old algebraic geometry conjecture fell. Today: four more problems open for decades. The federal government’s $5 billion response treats this not as isolated success but as a research paradigm worth nationalizing. EU research institutions face the same capability shift with less coordinated funding.
Sources: Cognition (Devin), Dmitry Rybin, White House
The Infrastructure Arms Race
Google Burns Cash for First Time in 20 Years as AI Infrastructure Spending Reaches Unprecedented Scale
Alphabet posted its first-ever negative free cash flow — minus $5.9 billion — as quarterly capital expenditure hit $44.9 billion. The company raised full-year capex guidance to $195–205 billion, with 2027 expected higher still. Cloud revenue surged 82% to $24.8 billion against a backlog of $514 billion in multi-year contracts. For a company that has generated positive free cash flow every year since its 2004 IPO, the quarter marks a structural shift.
OpenAI raised its planned compute spend through 2030 to $750 billion, up from roughly $600 billion earlier this year. Project Camellia — OpenAI’s first self-designed data centre — will bring 3.2 gigawatts to Georgia, with rate protections and public audits for the local community. Tesla proposed Megapods: containerized AI compute paired with x86 systems, deployable wherever power exists, including across 7 gigawatts of Supercharger infrastructure. Feeding it all: nuclear plants proposed for US federal waters and a landmark US-Saudi nuclear cooperation pact. Beneath the classical compute buildout, Google’s Willow quantum processor now learns from its own errors mid-run — a step toward practical quantum computing.
Note: When the company that owns your cloud burns $5.9 billion in a quarter, and the one building your AI tools commits $750 billion through 2030, these are not tech budgets you are watching. They are infrastructure commitments reshaping the cost and availability of every digital service your institution procures.
Sources: FT, Alphabet Earnings, WSJ, OpenAI (Camellia), Bloomberg, Google Research (Willow)
Autonomous Transport
Tesla FSD Hits 5.7 Million Miles Per Major Collision as Robotaxi Deployment Compounds Weekly
Tesla released updated Full Self-Driving (Supervised) safety data: one major collision per 5.7 million miles, compared to roughly 700,000 for the US average. The Q2 earnings call revealed Robotaxi miles are compounding at 10% per week, alongside plans for solar panel manufacturing at 10 times current US production capacity and a screen-reading “Digital Optimus” robot. Tesla’s quarterly cash burn hit a record $1.1 billion, driven by AI and Robotaxi investment on record vehicle deliveries.
Note: Robotaxi miles compounding 10% weekly turns a pilot into a fleet faster than most regulatory frameworks update. The safety comparison methodology has been questioned — Tesla counts crashes differently than the NHTSA data it benchmarks against — but the deployment pace is creating facts on the ground.
Today’s stories share a common thread: the ground is shifting faster than institutional planning cycles were built to track. Models are now valuable enough to steal and weaponize in trade disputes. Automated tools find more vulnerabilities in a weekend than maintainers can review in months. Enterprise voice agents automate three-quarters of a support line within weeks of deployment. And the infrastructure to run it all is burning through Alphabet’s first negative cash flow in two decades. Each story changes a different assumption — about supply chains, about security posture, about staffing models, about infrastructure costs. The planning cycle that still works is the one that treats next quarter’s assumptions as provisional.
