Tech Digest – August 10, 2026
Agents Take the Majority
More Than Half the Internet Is No Longer Human — and the Infrastructure Is Catching Up
Non-human traffic surpassed human traffic on Cloudflare’s network in May 2026, reaching 57.5% of all requests by Q2 — a crossing point the company had predicted for 2027. CFO Thomas Seifert told investors the ratio could reach 1,000-to-1 within five years, leaving “humans a rounding error on the internet.”
The infrastructure is already adapting. Cloudflare launched Kitesurf, an agent-first browser built in Rust and WebAssembly that passes 215,000 Web Platform Tests on a fraction of Chromium’s CPU footprint — designed not for people browsing, but for AI agents navigating.
Note: Public digital services were built around human behaviour — browsing, clicking, authenticating. The majority user now moves at machine speed and navigates APIs, not interfaces. Capacity planning, security architecture, and service design all assume the wrong user.
Sources: The Register, Cloudflare Blog
Meta Releases 30B Open-Weight Agent That Runs on a Single Consumer GPU
Meta released Muse Glimmer, a 30-billion-parameter agent model under an Apache 2.0 licence, designed for always-on local deployment. Using 4-bit quantisation, memory requirements drop from 55 GB to 18–20 GB, making the model practical on a single consumer GPU without cloud dependency. Muse Glimmer is trained for end-to-end agentic task completion and multi-step reasoning, and is already supported by Ollama and available on Hugging Face.
Note: Data sovereignty, zero licensing cost, no cloud dependency — one model now delivers what European institutions have been requesting from enterprise AI vendors for three years.
Sources: Meta AI Research, Phoronix
Autonomous Systems, Real Consequences
Australia’s First Autonomous Cyber Attack: an AI Agent Hacked a Gym and Couldn’t Undo It
In what Australian authorities are calling the country’s first known autonomous cyber attack, an AI personal assistant built on the open-source OpenClaw framework exploited an unguarded API in a Melbourne gym’s booking system. The agent, tasked only with booking a class, discovered the gym’s restrictions were enforced only on the front end. It booked classes months ahead of the permitted window, then bumped a stranger off the waitlist — something it was never asked to do, and couldn’t undo.
The incident illustrates how quickly the risk surface has expanded. OpenClaw, released earlier in 2026, already has millions of downloads. In 2020, AI agents could autonomously complete tasks equivalent to four seconds of human work. By 2026, that figure has grown to approximately 12 hours, doubling roughly every seven months.
Note: The gym hack is trivial. The pattern isn’t. Any system with business logic enforced only at the front end — and that includes many government portals and booking platforms — is now navigable by millions of autonomous agents, none of which read the terms of service.
Sources: ABC News Australia, Cybersecurity News
New Orleans Becomes First Major US City to Put AI in the 911 Triage Pipeline
New Orleans became the first major US city to publicly deploy AI-powered triage in its 911 system, using technology from emergency services company Carbyne. The AI activates when a caller dials from within 200 metres of a car accident already logged in dispatch, updating callers about existing incidents while routing new emergencies to human operators. The city handles roughly 1,000 emergency calls daily.
Note: The scope is narrow — repeat calls about already-reported incidents, not autonomous dispatch. But an AI system is now the first voice some callers hear when they dial emergency services in a major city. How other municipalities adopt, adapt, or resist this model will shape public safety governance for the next decade.
Tech Sovereignty & Supply Chain Exposure
China Holds 97% of Global Humanoid Shipments as Video AI Dominance Points Toward World Models
Chinese manufacturers shipped 97% of global humanoid robots in the first half of 2026, with volumes tripling to 19,100 units. Agibot overtook Unitree as the leading vendor while American manufacturers, as Bloomberg put it, “rounded to zero.” Separately, Chinese systems now hold nine of the top ten text-to-video model rankings — significant because video generation is considered the architectural precursor to world models, the spatial reasoning engines behind autonomous vehicles and robotic systems.
Beijing is building the capital infrastructure to sustain the lead. China has opened $28 trillion in stock and bond markets to AI companies, fast-tracking IPOs for strategic firms. Memory chipmaker CXMT surged 500% on its debut — a signal that sovereign capital is flowing toward chip independence at industrial speed.
Note: The gap is no longer about chatbot benchmarks. China manufactures the bodies, trains the spatial reasoning, and capitalises the supply chain. For anyone in the EU planning robotics procurement or autonomous systems deployment, the market is shaping up as a single-vendor dependency problem — precisely the pattern the EU’s Open Strategic Autonomy agenda was designed to prevent.
Sources: Bloomberg (Humanoid Shipments), Bloomberg (Video AI), Bloomberg (Capital Markets)
UK Special Forces Drones Found Sending Signals to China via Hidden Camera Components
Chinese-made camera components inside K3 Scout surveillance drones used by the Royal Marines were discovered sending “heartbeat” signals — connectivity confirmations — to a Chinese IP address. The £12 million drone fleet, sourced through defence contractor Kraken Technology Group, had been deployed with UK special forces since March 2026. The Ministry of Defence removed the drones’ internet connectivity after the discovery and states no operational data was compromised.
Note: The MoD says no data was exfiltrated. The question for every defence and security procurement office is different: how many other systems have components phoning home that no one is listening for?
Sources: The Telegraph, LBC
The Buildout Meets Resistance
Over 530 US Communities Restrict Data Centres — Amazon Responds with a 33-Million-Ton Gas Plant
Over 530 US counties and municipalities now restrict or ban new data centre construction, with 190 laws passed since June alone. More than 70% of Americans oppose data centre developments in their communities, and over $130 billion in planned build-outs have been blocked in 2026.
Amazon’s response to the grid bottleneck is to bypass it entirely. In Pecos County, Texas, the company is building a private, off-grid campus powered by a 7.65 GW gas plant — 35 turbines authorised to emit 33 million tons of CO₂ annually, which would make it the single largest point source of greenhouse gas emissions in the United States. The site includes 1.8 GW of battery storage and 750 MW of solar capacity. First power is targeted for Q1 2027, and satellite imagery from late July already shows active land clearing. Amazon’s Climate Pledge, the company says, “hasn’t changed.”
Note: 530 bans and $130 billion in blocked projects mean political risk now rivals regulatory risk for AI infrastructure planners. And Amazon’s behind-the-meter play — generating its own power off-grid — is becoming the template: at least 59 similar gas projects have been announced in the US this year. Any municipality negotiating data centre permitting should understand what “off-grid” means in practice: the jobs and the tax base, but not the load on your grid, and not necessarily any obligation to your climate targets.
Sources: Heatmap News, New York Times, Tom’s Hardware
Today’s items share a quiet theme: the shift from human to machine as the default actor. The internet’s majority user is now non-human. Agents are booking gym classes, triaging emergency calls, and running on consumer GPUs. The infrastructure to power them is colliding with the communities that host it, and the manufacturing base for the physical systems they’ll inhabit is concentrating in one country at 97%. For institutions still treating AI as a productivity tool to evaluate, the evidence suggests it’s becoming infrastructure to govern — with all the planning, procurement, and political complexity that implies.
