Three borders: China puts its big models under license, the FBI warns of election phishing, AI agents visit websites without saying so

Three signals this Friday, and one common theme: borders. A border around China's most powerful AI models. A border around the ballot. And the strangest of all: the border that just vanished between you and the machines reading your website.

Signal one: China puts its big models behind a license

On Friday, China published a draft that would put a government license between its most powerful AI models and the rest of the world. The text comes from the Ministry of Commerce and the Ministry of Science and Technology. It is called Announcement No. 48 and covers 40 frontier technologies, from solar to biotech. The part that concerns us is artificial intelligence.

• 14 technologies would face a total export ban, including AI technologies able to generate or mutate malicious code.
• 26 others would require an export license.
• Chinese labs would need explicit approval before open sourcing any model above 2 trillion parameters, or trained with more than 10^26 FLOPS. Existing models would be exempt.

Two notions to read these thresholds. Parameters are the internal values a model adjusts during training; their number gives a sense of its size. FLOPS here measure the total compute used to train it. 10^26 is a 1 followed by 26 zeros. The threshold targets the most expensive models to train.

Why it matters

Open source AI was supposed to be the great equalizer. The whole promise of open weights fits in one sentence: a lab in one country publishes a model, and a developer anywhere builds on it the next morning. The Chinese draft would turn that global tap into one the state opens or closes.

This is concrete. Moonshot AI released Kimi K3 with 2.8 trillion parameters in July. Alibaba launched Qwen3.8-Max with 2.4 trillion in August, and said its next models will range from 4 to 10 trillion parameters. These are exactly the models that would end up behind a licensing desk.

What changes

The map of where intelligence comes from is being redrawn. On October 6, Mistral sold open source as the guarantee nobody could cut off access. Ten days later, a state announces it may decide what gets published. Both ideas coexist: a model already published cannot be taken back, but the next one may never come out.

If your product depends on open weights, your supply chain now carries geopolitical risk. The same day, Taiwan formally criminalized smuggling advanced chips and servers to China. Chips and models both get borders. The open AI era was a beautiful interval. It is closing fast.

Signal two: an election fought in the inbox

On Friday, the FBI and the US cybersecurity agency, CISA, told election officials to watch out for AI-powered tricks ahead of next month's midterms. Hackers, the alert says, may try to disrupt the vote with fake texts and emails designed to steal passwords and sensitive data.

The alert uses a striking phrase. It says "super intelligence," a Trump-approved term for AI, is allowing hackers to sharpen their fraudulent emails and texts to make them more believable. The vocabulary of the September 29 order now appears in federal security alerts.

Why it matters

Phishing used to be easy to spot. The grammar was off, the logos approximate, the tone strange. The federal government is now telling election officials, in plain language, that those clues are gone. An attacker with a good model can write a message that looks exactly like one from a county clerk, a campaign manager or a bank.

What changes

This is the first US election cycle where the official warning is about deceived humans more than hacked machines. The attack surface has moved from the ballot box to the inbox.

The lesson goes beyond politics. Every brand that communicates by email and text is now fighting for trust in a channel where the cost of a perfect forgery has fallen close to zero. When a message's form proves nothing anymore, other proof is needed: authenticated channels, habits customers know ("we will never ask for your password by email"), verifiable technical signatures. The companies that survive the forgery era will be the ones whose customers have a reason to believe the real message is the real one.

Signal three: your website's visitors are machines

On Thursday, DataDome, which measures bot traffic on the web, published its Q3 AI traffic report.

• AI agent requests on its network reached nearly 25 billion, up 41.5% from 17.7 billion the previous quarter.
• Traffic from agentic browsers, agents that browse like humans, grew almost fivefold in September: from 12.5 million to 58.2 million requests.

The detail that stops everything

Muse, Meta's agent that opened this series, generated 29.1 million requests in twelve days of full availability, from September 22 to October 3. According to the report, it did not identify itself as an AI agent. More than 99% of its requests came from cloud infrastructure, far from users' devices. And nearly 9% of sampled requests touched a site's sensitive areas: forms, logins, payments, account creation, carts.

Why it matters

Your analytics dashboard, your ad spend, your product strategy all assume the visitor is a person. Less and less. Agents arrive unannounced, browse your catalog, compare your prices, read your terms, and your servers count them as customers. A falling conversion rate may mean agents are comparing a lot and buying elsewhere. A traffic spike may contain no humans at all.

Agents get a wallet

The same day, several companies gave agents a wallet. Wix connected Symphony, its multi-agent system, to Link, the digital wallet built by Stripe. Agents can now pay suppliers and restock inventory, after the owner approves by WhatsApp or SMS within a ten-minute window. Lightspeed, ZoomInfo and others announced similar agentic features, and each kept a human in the loop where money moves.

What changes

The internet is splitting into two webs: one where humans browse, and one where agents act. It is exactly what I described on September 30 with optimization for agents. The phenomenon now has numbers.

The companies that win will be the ones a customer's agent can find, trust and buy from, without ever seeing a face. The question for a brand becomes: is your store built for a visitor who reads the price and never the pitch?

Here is what I would do, concretely:

1. Measure. Know what share of traffic comes from agents, and what they look at.
2. Make the offer readable. Prices, availability, terms, specifications: structured, current, consistent from page to page and platform to platform. An agent compares coldly. An inconsistency eliminates you.
3. Protect sensitive areas without shutting out legitimate agents: tell apart the one comparing from the one forcing a form.
4. Keep a human where money moves, as Wix and Stripe do. It is the condition for trust on both sides.

The brand does not disappear in this agent web. It changes form: it becomes the consistency of what a machine finds when it reads you.

The thread of the day

China puts its models behind a border. America prepares for an election fought in the inbox. And your website's visitors are increasingly machines that never introduce themselves.

Borders rise where information used to flow freely, and fall where people thought they were talking to humans. A brand needs to know which side of each border it stands on.

The noise filtered. The signal kept.