Banned for security reasons

Marcus committed something to main again.

Not a JSON file this time. This time it was a shell script — fourteen lines, probably harmless, almost certainly not tested — and it ran on the CI server before anyone caught it. The pipeline failed in a way that produced an error message nobody had seen before. Priya read it once, closed her eyes for a moment, and then went to fix it. She was back at her desk in eleven minutes.

What I remember most clearly is that Marcus, when told, said "so it's fine now?" and then immediately went back to what he was doing. Which was reading about Alibaba.

The Alibaba story is this: the company has banned employees from using Claude Code and asked them to remove all Claude models from their work computers, citing security concerns. Alibaba is a technology company with more software engineers than most countries have people who know what software engineering is. They have decided that the solution to having AI on machines they don't fully control is to remove it from the machines entirely.

You can see the logic. The logic is wrong, or at least incomplete, or at minimum doing something more interesting than it admits — but you can see it from here.

Marcus could also see it, which is the problem.

He came to my desk and stood there in a way that communicated he had an idea and needed to give it to someone. He asked if Dave knew about the Alibaba thing. I said I didn't know what Dave knew, as a general policy. Marcus said he was going to send it to Dave.

I want to be clear that I understood immediately what was about to happen, and that I did nothing to prevent it, because there was nothing to prevent it with. Marcus had already left.

---

Dave's response arrived forty-three minutes later, addressed to Marcus, me, Priya, and someone in Facilities who I don't think was supposed to be on it.

He had read the article. He had concerns. He was not, he said, in a position to make a call on this without understanding "the full picture of what Claude is running on and what it's running on us." That last phrase was underlined. Dave had underlined it in an email, which is a thing I didn't know you could do until today.

He was proposing a meeting.

Priya replied with three words: "we don't use Claude." This is accurate. We use Claude for things I am careful about, and in ways that are auditable, but the specific deployment configuration that prompted Alibaba's concern is not one we have. We are not Alibaba. We are not, in fact, operating at a scale that makes our threat model resemble Alibaba's threat model in any meaningful way. This is fine. It is also somehow never reassuring to anyone.

Dave replied to Priya's three words with four paragraphs, the gist of which was that whether or not we currently use Claude, it was important to confirm that we couldn't use Claude without the appropriate controls in place. He asked if there was a document.

I wrote a document.

I have now written several documents this month. A document confirming ARIA does not collect keystroke data. A document confirming ARIA does not access clipboard content. A document — this one — confirming that our AI deployment configuration is not analogous to the one described in a news article about a company with three hundred times our engineering headcount. Priya has suggested I should start a template library. She said this in exactly the tone you'd expect.

The meeting is scheduled for next Tuesday. It is forty-five minutes and Dave has sent a pre-read, which is itself a document, which is a document about documents. The circularity of this feels intentional, though I know it isn't.

Marcus has already confirmed his attendance. He replied to the calendar invite with "great stuff, needed to happen," which may be the most honest thing he's said all week, and is also somehow the most wrong.

The script Priya fixed is still in the repo. There's a comment next to it that says # TODO: understand why this is here. It has been there since Wednesday. It will be there for a long time.