OpenAI's Bugged Boardroom and Anthropic Ditching Git

OpenAI found a recording device hidden in one of its main meeting rooms. An employee planted it, camouflaged with toilet paper.

On the calmer end of the news, Anthropic replaced Git with its own in-house version control. Its coding agents were merging more than 30,000 code changes a day. Two very different stories about machines in the office.

The recording device in an OpenAI meeting room

Here is what OpenAI says happened. An employee planted a small recording device behind a television in a main meeting room.

It was wrapped in toilet paper so nobody would look twice. It ran for some period of time, feeding what it captured to outside parties.

OpenAI has not said how long it was live. That detail is doing a lot of work by being missing.

The part everyone is talking about is how they caught him. OpenAI fed the device's data into their own internal models. The models helped identify who built the scheme and who was responsible.

The employee was fired on the spot and now faces federal charges.

You have to admire the product demo. A company that sells intelligence used its own intelligence to find the mole in its boardroom. The enterprise sales deck more or less writes itself.

It is also a message to anyone else inside the building with similar plans. Not a loud one. Just a quiet note that the house has very good eyes.

I do not think this is a one-off. AI labs now attract the full menu of attention.

Foreign intelligence services. Activists who think the work threatens humanity. Competitors, and people who just want screenshots.

The hiring bar at these companies is about to get much weirder. Expect more background checks and fewer benchmark scores.

There is a second, less comfortable thread here. If labs point their models at a security incident, they can watch staff activity too.

Anthropic in particular has spent years warning about AI-powered surveillance. Expecting your employer not to monitor you here is optimistic.

For you, running a normal business, two practical things come out of this. First, physical security of meeting rooms is cheap. A recording of your pricing talks is not.

Second, and this one matters daily. Whatever your staff types into a chatbot is stored somewhere.

It can be requested in a lawsuit. It can be read during an investigation. That is one reason governments are moving to open-weight models on their own hardware.

You do not need to go full paranoid. You do need a written rule about what goes into which tool. And a provider whose data policy you have actually read.

Why Anthropic replaced Git with its own version control

Now the story that will affect more businesses, even though it sounds like plumbing. Anthropic published how it built a new version control system. It moved its entire main codebase off Git over five months.

Version control tracks code changes, who made them and in what order. Git is the standard. GitHub is where most teams keep it.

Anthropic decided the standard no longer fits what they do.

The numbers explain why. In May, they merged about 2,500 code changes a day. By early September, more than 30,000 a day.

That is roughly twelve times more in four months. The growth comes from agents writing code, not twelve times more engineers.

Their own phrasing is the useful bit. They were bottlenecked by developer infrastructure built for human scale rather than agent scale.

Every tool in a normal software team assumes a human types and a human reviews. Remove the typing human and the rest of the pipeline starts groaning.

I would not read this as "GitHub is finished". Anthropic is an extreme case with extreme resources. But the shape of the problem is coming for everyone who puts AI agents into real work.

Here is what it looks like outside a coding team. You give an agent your support inbox and it drafts 200 replies a day. Your one approver is now the bottleneck.

Or an agent takes over reporting. Now it produces 40 documents a week that nobody reads. The work got cheap. The checking did not.

So before you add agents anywhere, look at the step after them:

• Who reviews the output, and how many items per hour can they handle?
• What happens to the queue when volume jumps ten times?
• Which outputs can go out without a human, and which absolutely cannot?
• Where do you store it all, and can anyone find anything a month later?

Answer those and you avoid the expensive version of this lesson. Ignore them and you buy a very fast machine for making work nobody reads.

The post quietly shows something else. Big labs will keep absorbing the tools they used to buy. If a lab depends on a piece of software, its agents can rebuild it.

If you sell a thin product sitting on top of a model, think about that. The defensible part is usually the messy stuff labs avoid.

Your industry's rules. Your customers' habits. The integration nobody enjoys building.

What I'd take from both

These two stories are closer than they look. In both cases, a lab pointed its own models at its own internal problems. One caught a spy. One rewrote how its code ships.

Your version of that is smaller and far less dramatic. Look at the processes that exist because a human used to do every step. Most were designed around human speed, and nobody has questioned them in years.

And if you run a company worth spying on, check behind the TV. It costs you ten seconds. The alternative apparently costs you a federal case.