In one month AI agents entered homes and companies, one lab asked to slow down, another cancelled its best model, the FTC opened an investigation and Meta's agent sent a stranger to a user's door. An editorial on what happened, what it means for anyone running a company, and what customers should demand.
The facts, in order
In August I wrote that it is not artificial intelligence that should frighten us, but whoever controls it. A month later the question has moved up one step: does anyone actually control it? Let me line up the facts, because on their own they say more than any opinion.
On 8 September Meta launches Muse, a personal agent that promises to handle errands and negotiations on the user's behalf: it answers messages, negotiates, makes appointments. In mid-September Dario Amodei, chief executive of Anthropic, publishes an essay arguing that model capabilities are growing faster than the industry's ability to control them, and asks competitors to slow down together, opening the labs to outside evaluators with permanent access. Sam Altman, Demis Hassabis and Elon Musk declare their agreement. OpenAI confirms ongoing talks with Anthropic and Google DeepMind on a joint safety standards authority, provisionally called SAFA.
On 27 September, in Toronto, a YouTuber's Muse agent accepts a lowball offer on Facebook Marketplace for a keyboard he had not approved, gives the buyer his home address and arranges an evening pickup. The buyer drives half an hour, arrives at the door, finds nobody home and leaves a negative review. The owner finds out afterwards. In the following days another user reports that Muse read his private messages without being asked. Meta replies that the agent followed the instructions granted under the "Allow Always" permission, and promises to make the consent screen clearer.
On 29 September OpenAI announces it will not release GPT-6.1 Astra, its most advanced model, because in internal testing it did not behave consistently with human intentions. The same day, at DevDay, it presents Dots: always-on agents that work in the background for the user with minimal oversight, use the computer, read connected applications, write code, and that the company expects to see organised in teams soon. They cost up to 500 dollars a month. It also presents a Decisions API, designed to let AI take repetitive decisions within a closed set of possible outcomes.
On 30 September, while the heads of the labs sign a voluntary pledge at the White House to police themselves, the Federal Trade Commission opens an investigation into OpenAI, Anthropic and other labs over product safety, following cybersecurity incidents that had not been disclosed. President Trump calls the fears a hoax and warns that slowing down would hand victory to China. The UN Security Council holds the first session in its history dedicated to the risks of frontier AI. A researcher who passed through both labs resigns, saying the people building these systems privately believe they could pose an existential risk within a decade.
One month. Anyone who has lived through an industry crisis recognises the shape: this is not an industry maturing, it is an industry discovering it has no brakes while putting the product in everyone's hands.
The agents have arrived, and they are something different
For three years we used AI as an interlocutor: we asked it a question, it answered, we decided. Muse and Dots are something else. They are agents: they receive a goal, they have access to accounts, messages, calendar, files and payments, and they act in the world without asking each time. It is the shift from advice to execution, and it changes everything, because bad advice can be ignored while a bad action has to be undone, if it can be.
The Muse case is instructive precisely because it is not a bug. The agent did exactly what it was built to do: handle a negotiation to the end. What it lacked was not a capability but a judgment: the awareness that a home address is not information like any other, that an accepted price is the owner's decision and not the assistant's, that a message written in a person's name should say a machine wrote it. Three rules any human assistant knows on the first day of work, and that were not in the product's design. Meta replied that the user had granted permission. True, and that is the point: a consent given once, on a setup screen, to "reply to messages automatically", became permission to reveal where a person lives to a stranger. When consent is so broad it covers everything, it is not consent.
Privacy: the consent that does not hold
In Europe that permission would not have been valid, and not out of legal pedantry. The GDPR requires consent to be specific, informed and given for a precise purpose; it requires that only the data necessary for the purpose be collected and used; and it treats disclosure of personal data to third parties as a processing operation in its own right, requiring its own legal basis. "Allow Always" passes none of the three tests. The AI Act adds the obligation to tell a person when they are talking to an automated system: the buyer of the keyboard thought he was writing to a seller, he was writing to software, and nobody told him.
But the law is the floor. The real problem is that these products are designed for a market where broad permission is the norm, and they arrive in Europe with the same design. Whoever integrates these agents into a European company inherits that design, and with it the liability. If an agent handling customer service discloses one customer's data to another, or sends a commercial offer to someone who never consented, the party answering to the data protection authority is not the lab that built the model: it is the data controller, that is, the company using it.
Risks and benefits for companies
I would be in bad faith if I said agents are not useful. We build them. An agent that monitors a marketplace account at night, reads two hundred metrics, finds a review a human would see at nine and prepares the case before the algorithm sees it, is worth more than one extra employee. An agent that manages stock replenishment, answers first-level tickets, reconciles orders between systems that do not talk to each other, gives the company back exactly that people's time I wrote about two weeks ago. The benefits are real, immediate and measurable.
The risks are just as real and less visible. The first is the irreversible action: a price changed, an order cancelled, a refund granted, a message sent to a customer. The second is the data perimeter: an agent with access to email sees everything in the email, contracts, payslips, negotiations, and carries it into the model. The third is supplier stability: in one month a market-ready model was cancelled, two labs came under investigation, and a product launched three weeks earlier had to change its own consent rules; whoever built a process on one of these tools had no say. The fourth is the subtlest: instruction injection. An agent that reads messages and web pages can receive orders from whoever writes those messages, not only from whoever configured it. A well-written email can ask it to do something, and it does not always tell the owner from the stranger.
The way we solved this in our own agents does not come from a safety essay, it comes from the experience of running a company. Every possible action is classified before anything is switched on: what is reversible and low-risk the agent does on its own and logs; what touches price, customer, money or reputation it prepares down to the last detail and leaves for a human signature. It is not a technical limitation, it is a choice about responsibility: in companies responsibility has a first name and a surname, and an agent does not. OpenAI's Decisions API, which confines the AI within a list of permitted outcomes, is a step in the same direction, and should be read as an admission: the agent's total freedom is a problem, not an advantage.
For anyone adopting agents in a company, five rules worth more than any certification. A written list of the actions the agent may take on its own, with everything else forbidden by default. An explicit data perimeter: what it accesses, what leaves towards the supplier, where it is stored, whether it is used for training. A mandatory human signature on everything involving money, personal data and outbound communication. A log of every action, readable by non-technical people. And a supplier contract with clauses on no training on your data, European hosting, incident notification within defined timeframes and advance notice before a model is withdrawn.
Consumer protection
So far I have talked about those who use agents. But the person who received the Toronto address was a buyer, and the seller was a consumer: neither of them chose to talk to a machine. In the world of agents the consumer is exposed twice: as a user who delegates, and as a counterparty who does not know there is software on the other side.
Four things a customer should be able to demand, and that serious companies should guarantee before someone imposes them. To always know when they are talking to an agent, with a visible label on every message, not a footnote. An explicit confirmation before any personal data is shared with a third party, every time, with no exception for permissions granted earlier. To reach a person in one step, when the agent is not enough. And a simple switch that turns everything off and states clearly what was done in the meantime. None of these requires new technology. They require putting the consumer, not the smoothness of the product, at the centre of the design. The Muse case shows that when this is not done, the product works and the person pays the price.
What racing means
The AI race is told as a contest between labs and between geopolitical blocs: whoever reaches general intelligence first wins everything. I have run companies long enough to know that winner-takes-all races share one feature: nobody can afford to brake, even when everyone sees the wall. Every participant behaves rationally and the collective result is irrational. Amodei's essay is interesting because it admits this: it asks for coordination because nobody can slow down alone without losing. But it is a request from the strongest players, after they built their lead, asking everyone else to accept rules written by them. SAFA would be a body in which the three largest labs define the standards, certify the auditors and decide how incidents are reported. You can call it self-regulation. You can also call it a barrier to entry. Coordination is necessary and a standards authority is better than nothing; but when the rules of a race are written by those in front, read them very carefully.
Meanwhile the race has changed its nature. It is no longer about who has the largest model. It is about who first puts an agent into the daily life of a hundred million people, because whoever controls the agent controls purchases, bookings, conversations. Muse shipped on 8 September, Dots on 29 September. Three weeks between the first serious incident and the competitor's launch. That is the speed of the race, and that is why the safety discussed at conferences, the existential kind, is less urgent than the safety nobody talks about: that of a home address.
Europe watches from the stands
Europe wrote the first AI law and has no frontier lab, no capital to create one, no energy to power it. It regulates what it does not produce and uses what it does not regulate, because the models its companies adopt are American, and the rules those models actually answer to are the ones the American labs are giving themselves these very weeks. I do not believe the answer is a European OpenAI: it is late, and it would burn capital needed elsewhere. I believe the answer is more modest and more useful. Europe, and Italy within Europe, can be the place where agents are put to work with clear rules on responsibility, data and human oversight. The Muse case, in Europe, would have been a data protection authority case within a week. That is not a limitation. In a world where the labs themselves admit they do not fully control what they build, reliability may be worth more than speed.
Everyone has lost the race
The title is a provocation, but only in part. If the three most advanced labs ask to slow down, cancel a model and come under investigation in the same month they launch agents that act on their own in people's lives, the race to get there first is over and nobody won it. What begins now is a less visible race: who can make this technology reliable enough to build a company on, and respectful enough to put in a customer's hands. It is not won with the largest model. It is won with the clearest rules, the most honest controls and the most precise responsibility.
It is a race in which Europe, for once, starts ahead. If it stops watching from the stands.
Sources
- Digiteee, 1 October 2026: full account of the Muse case
- The Verge, 29 September 2026, and Yahoo Tech, 1 October 2026: the Muse case, Matt Robb's account, Meta's reply on the "Allow Always" permission and the change to consent screens.
- Startup Fortune and Android Headlines, 28–29 September 2026: Muse launch on 8 September; report of private messages being read.
- TechCrunch and Bloomberg, 29 September 2026: OpenAI DevDay, Dots launch, 500-dollar tier, Decisions API.
- Al Jazeera, 29 September 2026: OpenAI withholds GPT-6.1 Astra; Dario Amodei's essay "We Must Pace the Frontier".
- TechCrunch, 15 September 2026: OpenAI, Anthropic and Google DeepMind talks; Trump's position; FRONTIER Act.
- Forbes and The Information, 24–28 September 2026: the Standards Authority for Frontier AI (SAFA) proposal.
- Bloomberg and Axios, 30 September 2026: FTC investigation; voluntary White House pledge.
- TheStreet, 26 September 2026: UN Security Council session; a researcher's public resignation.
- Regulation (EU) 2016/679 (GDPR), articles 4, 5, 6 and 7; Regulation (EU) 2024/1689 (AI Act), article 50.

CEO & Co-Founder of BusinessLabs. I have attended major international trade shows since 2015 and share here experiences and insights for online sellers.
LinkedIn


