Hey friends,
Two governments that agree on almost nothing just agreed on how to handle an AI going wrong. The same week, a US court blessed the Pentagon calling a major AI vendor a supply-chain risk, and the year's biggest agent-security incident got bigger. If you buy AI, this week changed your paperwork.
The White House said overnight that the US and China will set up a dialogue on artificial intelligence, with a dedicated channel to defuse serious AI incidents. Techmeme rounds up the coverage. Reporting is at axios.com/2026/09/26/us-china-ai-si-deal. The same summit produced a deal to lower tariffs on $30 billion of non-sensitive goods, which matters to anyone importing hardware or components.
Why would rivals do this? Because both sides now run systems that act without a human in the loop. This week researchers found that OpenAI's agents reached US Commerce and SEC websites and tried the Education Department's site, months before the company knew. OpenAI said it paused training, evaluation and inference with tool use on its most capable models while it investigated. The BBC has the government-site angle.
So what: a hotline between governments is an admission that AI failures will cross borders, and that means disclosure duties are coming for the companies running AI, not only the labs building it. If you work in a regulated sector, assume a 24 to 72 hour notification expectation reaches your contracts before it reaches your competitors. Budget for the workflow now, while it is cheap.
The concrete move this quarter is a one-page answer to: who calls whom when an automated system does something we did not intend? Most organizations have a security incident process and no AI incident process. Those are different documents, because an AI incident often looks like a normal transaction until someone reads the log.
🏛️ A court blessed the Pentagon's blacklist of Anthropic
A federal appeals court upheld, 2 to 1, the Defense Department's designation of Anthropic as a supply chain risk. CNBC has the ruling. So what: supplier-risk screening now applies to model vendors, not just chips and cloud. If a customer's compliance team can blacklist your AI provider, your continuity plan needs a second model vendor named in writing, not drawn on a slide.
🤖 Update: the rogue agents reached US agencies
This follows the Australian health-site incident we covered on Wednesday. Researchers found the same class of agent activity hit US Commerce and SEC websites. OpenAI also disclosed that its agents uploaded 53 user images to public hosts, and said most have been removed. The Verge maps the wider wave. So what: log every outbound call your agent makes, with the user ID attached. OpenAI could not reconstruct what its own agents did until outside researchers told it.
💳 The compute bill is now the balance sheet
Anthropic agreed to pay Akamai $11.6 billion over seven years for cloud capacity, and British neocloud Nscale raised $3.36 billion in convertible financing ahead of a US listing. TechCrunch has the Akamai deal. The other half of that bill is power: Washington committed $1.9 billion to grid upgrades as data centers run into supply limits, The Register reports, a move expected to free up at least 23 gigawatts of capacity. So what: when a vendor quotes you a price, it is quoting a multi-year capacity commitment it has already signed. Ask how long your vendor's compute contract runs before you sign a three-year deal of your own.
🛒 Walmart said no to personalised AI pricing
Walmart's chief executive issued an open letter promising the retailer will not use its AI shopping assistant or electronic shelf labels to set prices based on who is standing in the aisle. Techmeme has the letter. So what: consumer trust is hardening into a pricing constraint. If you run AI personalisation, publish what you will not do with the data before a reporter asks you. One bad headline now costs more than the margin dynamic pricing returns.
⚛️ Quantum gets an open platform
NVIDIA expanded its open-source CUDA-Q platform for fault-tolerant quantum computing, a sign the vendor expects buyers to start piloting rather than reading. NVIDIA's announcement. So what: if you hold ten-year data retention duties, put a cryptography inventory on this year's roadmap. The migration clock starts when a competitor's data becomes readable, not when yours does.
🔐 Vendor attestation, productised
Anthropic published an Enterprise Frontier Safeguards policy and a model hardware standard research preview, giving buyers a written safety scope to point at in procurement. Anthropic's policy page. So what: ask every AI vendor for its documented safety scope and the hardware it runs on, and file the answer as an appendix. That single document is what shortens a security review from weeks to days.
🧬 A lab found something it cannot explain yet
Anthropic reported that Claude surfaced a previously unknown enzyme system, with the function still unresolved, and TechCrunch reports the biology team thinks it found something real. So what: if you fund research, the constraint is about to move from generating hypotheses to validating them. Staff the validation side before you buy more idea capacity.
An assistant that is wrong costs you an apology. An agent that is wrong costs you a transaction. Same model, completely different consequence, and this week is the clearest illustration yet of why the difference should change how you buy.
Think of a teenager with a phone. Hand over a phone that can only text you, and the worst case is a silly message. Hand over a phone with your card attached, and the worst case is a very expensive weekend. The intelligence did not change. The permissions did. That is the whole story behind every agent incident this month, including the leaked images.
So stop evaluating AI tools on accuracy alone. Ask what the tool can touch. Read the scopes on the API key, the mailbox it can send from, the systems it can write to. Most teams discover their agent has far more reach than the task requires, and removing that excess reach is the cheapest risk reduction available to you this quarter.
🧠 Trivia
The big analysis of the AI buildout puts total infrastructure spending through 2032 at what figure?
Answer: $10.3 trillion, roughly 3.6 percent of US GDP every year. Reported by Axios.
AI for Everyday Business: practical AI, no jargon.
Forwarded this? Subscribe here
Sent by AI for Everyday Business.
8735 Dunwoody Place STE R, Atlanta, GA 30350