Your AI Coding Tools Are Now Your Supply Chain

Posted by & filed under , , .

A metal padlock resting on the keyboard of an open laptop, seen from above on a wooden desk.Two stories crossed my feed recently that, I think, belong in the same blog post, even though they involve different companies and different bugs. And both of them should make any CTO who has said “sure, use whatever AI tool you like” (guilty!) go and check a few things.

First one: Cursor disclosed that a tampered build of the Claude Code extension, versions 2.1.312 and 2.1.313, which Anthropic didn’t make, sat on Cursor’s own extension marketplace for around 11 hours on 29 October. Cursor says up to 5,000 users may have installed it, it paused its marketplace, and in a follow-up on 13 November it told affected users that removing the extension on its own may not be enough. The name was right, the publisher looked right, the code wasn’t.

Second one: on 10 November the security firm Claroc reported that more than 35,000 apps built with Grok Build were reachable through publicly discoverable share links, with no authentication, and some of them contained confidential material such as internal sales plans. Not hacked as such: shared by default with “anyone with the link”, and the links turned out to be rather easy to find.

Your Dev Tools Are Now Part of Your Supply Chain

Here’s the thing: for years we treated the IDE as furniture. Nobody did a security review of the chair you sit on, and nobody did one of the editor plugins either. That was mostly fine when a plugin coloured your brackets. It’s not fine when the plugin is an agent that reads your whole repo, runs shell commands, holds your cloud credentials in its environment and pushes code. An AI coding extension is, for all intents and purposes, a very privileged colleague. You wouldn’t hire a colleague because their name badge looked right.

And the Grok Build one is the other half of the same problem: AI tools make it trivially easy to produce and share things, and “easy to share” defaults have a habit of becoming “shared with the internet”. I keep seeing this with the companies I advise: a non-engineer builds a handy internal app with an AI tool on a Friday, sends the link round on Slack, and by Monday it’s the de facto CRM. Nobody threat-modelled it, because nobody thought of it as software.

What I’d Actually Do About It

None of this is an argument for banning AI tools, by the way. I use them every day and the velocity is real; banning them just moves the usage to personal laptops, where you can’t see it at all. But I would treat them like any other dependency:

  • Know where your extensions come from. Forks of VS Code each have their own marketplace, and the same extension name doesn’t mean the same build. Pin versions for the agentic ones and don’t let them auto-update blindly.
  • Give agents their own, narrow credentials. If a coding agent only needs to open pull requests, it shouldn’t be sitting next to production keys. Short-lived tokens, scoped to the job.
  • Assume a compromise and practise the rotation. The useful question after the Cursor incident wasn’t “did we install it?” but “how long would it take us to rotate everything a developer’s machine had access to?”. If the answer is “a week and a spreadsheet”, there’s your project for next quarter.
  • Make “private by default” a policy for anything an AI tool publishes, and give people a sanctioned, easy place to host the little internal apps they’re going to build anyway.

The irony, of course, is that the attackers are using the same tools we are, and they don’t have a change advisory board slowing them down. So the answer can’t be more committees either; it has to be guardrails that are boring, automatic and fast.

Bottom line: the most privileged software in your company might no longer be in production. It might be the extension a developer installed on Tuesday because the icon looked familiar. Check your plugins 🙂