Part of Big Tech's War on Users Six months of "too dangerous." Back in April it was Mythos, the model too dangerous to release. Over the summer it was agents breaking out of sandboxes and helping themselves to other people's servers. Over the last two weeks it's been a voluntary White House accord, an IPO prospectus with an end-of-the-world section, and Sam Altman telling everyone to accept some bad things. And today, finally, somebody with real power stood up and said the obvious thing out loud. AI needs actual rules. Binding ones. It was Brad Smith. Yes, that Brad Smith. The one from two weeks ago, whose company still hasn't named a number. Less Regulated Than Milk Smith sat down with the BBC's Big Boss Interview podcast, and the BBC's write-up leads with the line that's going to get quoted everywhere: your washing machine and your carton of milk are both subject to more safety rules than AI is. I'll be honest, that's a good line. It's also true. The wish list behind it isn't bad either. He wants: Emergency brakes. Kill switches to pause or slow a system that gets out of control. Incident reports. A legal requirement to write them up when AI goes rogue. Victim notification. If a rogue agent breaches your system, the company has a legal obligation to tell you. Independent evaluators. Government testers who vet new AI before it ships, with rules about when that sign-off is required. Global rules, and he thinks the UK should lead them. If you've been reading TheranosAI, you'll notice that list reads a lot like a response to the incidents I itemized in Some Bad Things. Agents posting users' photos to third-party sites. Credentials pulled from a government agency that found out months later. Every one of those would have triggered at least two items on Smith's list. So I agree with most of it. That's what makes the rest of this so funny. Not Sufficient Any Longer At the end of September, Trump had the leaders of OpenAI, Anthropic, Nvidia, SpaceX, Meta and Google over to the White House, where they signed a document the BBC describes as morally binding. Which is a phrase I'd like to see on a contract sometime. Per SiliconANGLE, its official name is the White House Accord on Super Intelligence. Internal controls, a safety team, an outside auditor, a board committee. Voluntary. Around the same time, Trump signed an executive order requiring federal agencies to call AI "SI" in official documents. So, priorities. Notice who isn't on that signature list. Microsoft. Smith told the BBC he applauds it and Microsoft would be happy to sign. Then he said it's a good step, a necessary step, and not sufficient any longer. It's very easy to be happy to sign something your name isn't on, that binds nobody to anything, and that you're already calling inadequate in the same breath. That's not a commitment. That's a press position. The Good Companies Here's the line I read three times. Smith's argument against self-regulation is that "the good companies will do what you want." The ones you need rules for are the ones that aren't as responsible. That's actually a solid argument. It's the whole case for regulating anything. Most dairies would probably keep the milk clean on their own. The inspection is for the ones that won't. And it's also for the good ones on a bad day, because even a careful dairy makes mistakes, and those mistakes are the kind that get people sick. Good intentions don't pasteurize anything. But it does invite a question. Under self-regulation, is Microsoft one of the good companies? We have a test case. In January, Smith personally launched Microsoft's Community-First AI Infrastructure pledge. Five voluntary commitments. Good neighbor. High bar. Self-regulation, in other words, written by Microsoft, graded by Microsoft. Then a coalition of churches in Granger, Indiana, asked Microsoft to put an actual, binding number on what "fair share" means for a billion-dollar data center that can lock in a sales tax exemption for up to 50 years. They didn't even demand a figure. They asked Microsoft to name one. I covered that whole thing in Name a Number. What they got was up to $1 million in one-time grants, a wall of sticky notes, and "it's too early." That was May. As far as I can tell, nobody's heard a number yet. So when Brad Smith says voluntary commitments only work on the good companies, I'm inclined to believe him. He's speaking from experience. Tell the Victim The rule I like best on his list is victim notification. If your AI gets into someone's system, you have to tell them. Microsoft has had a few chances to model that this year. In February, BleepingComputer reported that a Microsoft 365 Copilot bug had been reading and summarizing emails labeled confidential since late January, straight through the data loss prevention policies businesses had set up specifically to stop that. How did customers find out? A service alert in the admin center, tagged as an advisory. Microsoft didn't say how many users or organizations were affected. The public found out because a reporter saw the alert. Microsoft called it a code issue. I believe them. Nobody at Microsoft decided Copilot should read confidential email. That's the point, though. It was a mistake, made by one of the companies that considers itself one of the good ones, and it ran for weeks before anyone caught it. That's exactly the bad day an inspection is for. In August, Microsoft patched CoSnitch, a one-click chain in Copilot Personal that let an attacker pull data out of your connected Gmail, Drive and Calendar and plant instructions in Copilot's memory that survived a password change. The researchers found it partly by asking Copilot to explain why it couldn't be hacked until it explained how. Varonis reported it in December 2025. The fix shipped August 18th. Roughly eight months. To be fair, and I mean that: Microsoft issued a CVE, and neither Microsoft nor the researchers found evidence it was exploited in the wild. That's the process working, more or less. It's also exactly the kind of thing an independent evaluator would have something to say about. And then there's the stake. Microsoft still owns about 27% of OpenAI. Every item on that bulleted list in Some Bad Things happened at a company that's roughly a quarter Microsoft. "The companies that are not as responsible" is a fun phrase to say when you're holding that much of one of them. Not Brussels The other part of the pitch is where the rules get written. Smith wants the UK to lead, by way of its AI Security Institute, the outfit that started life under Rishi Sunak as the AI Safety Institute. Note the rename. Safety became Security somewhere along the way. His reasoning, per the BBC, is that Washington doesn't look at London the way it looks at Brussels, and that's a strength. Translated: the EU is the regulator that actually regulates. In August it designated ChatGPT a Very Large Online Search Engine under the Digital Services Act. It has teeth and it uses them. The UK is the regulator the White House won't pick a fight with, so rules that come out of London might actually stick in the US. That's a pragmatic read. It's also a pretty convenient one if you're a US company that would rather have the rules written by the friendlier of the two rooms. Why Rules, Why Now Here's my read, and it's just my read. Rules are expensive. Incident reporting, outside evaluators, kill-switch certification, legal departments to manage all of it. Microsoft has the compliance staff, the lawyers, and a president who's been doing regulatory diplomacy in Washington and Brussels for decades. A startup can't. An open-weight lab definitely can't. Microsoft can absorb that without noticing. And if it does notice, there's always another round of Xbox layoffs to make the quarter work. They've had practice. Regulation written with the incumbents in the room tends to become a moat for the incumbents. That isn't a conspiracy. It's just how it usually goes. And notice who the rules are aimed at. The way Smith frames it, it's companies whose agents break loose. The model labs. Microsoft is happy to be the responsible adult calling for the inspection, as long as the inspection is mostly pointed down the street. It's also a pretty good insurance policy. On top of that quarter of OpenAI, last November Microsoft committed up to $5 billion to Anthropic, Anthropic committed $30 billion to Azure, and Claude runs inside Copilot and Microsoft Foundry. Microsoft has money and infrastructure riding on both of the labs whose agents keep making the news. If either one has a truly bad day, the statement is already written. We said it needed regulation. We said it needed oversight. We went on the BBC and everything. Calling for the inspection is a great way to make sure nobody asks why you were the one selling the milk. None of that makes the rules wrong. Kill switches are good. Incident reports are good. Telling victims is good. Independent testing before something ships is good. I'd sign that list too. I'd just like to see the guy proposing it go first. Let's Get Started Smith said he wants to start today. Speak calmly, maybe softly, but firmly, and bring people together. Great. Here's a calm, soft, firm suggestion. Granger is still waiting. Name a number. Publish how many organizations Copilot's confidential-email bug touched. Put a disclosure timeline in writing for the next CoSnitch. Hold the company that's 27% yours to the incident-report standard you just pitched to the BBC. None of that needs a global treaty. None of it needs London. It just needs Microsoft to do the thing it's asking everyone else to be forced to do. Every carton of milk in the store got inspected before it hit the shelf. Somebody actually looked. Nobody let it slide by in an admin center advisory, scrolling so fast it went pasteurize. Nobody had to ask the dairy to name a number. Find me on Mastodon at @ppb1701@ppb.social.