Skip to main content

Security: Cursor explains how a lost key at a subcontractor trapped Claude Code


Two weeks ago, I told you how a fake Claude Code extension had trapped up to 5,000 Cursor users, the code editor designed for working with AI. At the time, one question remained unanswered: how had the hackers managed to slip their tampered version into Cursor's store? On Thursday, November 13, Cursor published the results of its investigation. And it's a real game of dominoes, where the first one falls at a subcontractor's and the last one onto your computer. Spoiler: nobody needed to hack anything in the Hollywood sense of the term, no hood, no green screen with scrolling text.

In an office, a long row of dominoes is falling, the first one pushed by a small key sitting on the table, and the last one about to topple onto an open laptop

A small key at the beginning, a trapped computer at the end

First domino: a key entrusted to a subcontractor

Cursor doesn't run all its AI models itself. It entrusts some of them to Fireworks AI, a company that rents out its computing power to run other people's models. To work together, Cursor had given Fireworks an access key to an online storage space, with read-only access. According to Cursor, it was Fireworks that allowed this key to leak. Classic: it's always the neighbor's fault. Fireworks, for its part, hasn't said anything publicly about this in the reports I've been able to read.

And in that storage space, there were “traces”. A trace is the logbook of a work session with Cursor's AI agent: what the developer asked for, the code the AI read and wrote, the commands it ran. These mostly dated from 2025, and came from people who hadn't enabled Cursor's privacy mode, “Privacy Mode”. Without this mode, Cursor is allowed to use these sessions to train its models. That was the case here.

Second domino: passwords forgotten in conversations

The problem is that some of these traces contained credentials in plain text. Access keys to online services, passwords, left exactly as they were in the session text. Cursor doesn't say how they got there: pasted into the conversation by a developer in a hurry, read by the agent in a file, who knows. But once they were in the log, they stayed there.

On a messy desk, an old notebook is open, covered in illegible scribbles, with a yellow sticky note stuck in the middle of the page, bearing a scribbled password, and a spilled cup of coffee beside it

The session logbook, with a forgotten password in the middle

The hacker recovered these credentials, and they opened other doors for him: internal Cursor tools, old code repositories that were no longer used in production, customer support data, emails, usage statistics. And above all, one of them allowed him to write to the extension store's cache, a sort of reserve where Cursor keeps copies of extensions to serve them faster. On October 29, the hacker deposited his tampered copies of Claude Code there, versions 2.1.312 and 2.1.313. They remained in place for about eleven hours.

Cursor insists on this point: Anthropic never published these versions, and “this was not a vulnerability in Anthropic's product”. The original extension repository, Open VSX, never hosted the malware either.

How many people were affected?

Cursor gives three figures, and be careful, they don't count the same people. Fewer than one user in two thousand had exposed traces. Fewer than one customer in ten thousand saw their access keys to online services leak. And at most 5,000 people could have received the fake extension on October 29. Cursor doesn't say how many users it has in total, so it's impossible to turn these fractions into a number of people.

The good news is everything that wasn't affected. People who had enabled Privacy Mode lost nothing. The Cursor application itself, its official updates, the keys used to sign its versions, the clients' code: none of that was reached, according to the investigation. And the rule is simple: “if you didn't receive a notification, your data wasn't in these traces”. Cursor warned the people concerned one by one, by email and in the application, starting on October 30.

On the repairs side, Cursor stopped running its models at Fireworks, changed all its internal passwords as a precaution, rebuilt its extension store and added a secret detector, a tool that automatically spots a password or key lying around where it shouldn't be. Last activity by the hacker detected: November 1, and it was blocked.

A large, thick steel safe, firmly closed, sits in an office, but a yellow sticky note is stuck to its door right next to the dial, with numbers scribbled on it

An armored safe, and the code on a sticky note

What you still have to do

If you installed or updated Claude Code from Cursor's store on October 29, to versions 2.1.312 or 2.1.313, follow the procedure in the October 29 security notice. Cursor warns that “uninstalling the extension may not be enough”. And if you're an administrator of a team and Cursor sent you a list of keys to change, change them all, not just the ones that seem important to you.

For everyone else, three reflexes. If you use Cursor for code that concerns only you or your company, go to its privacy settings and check that Privacy Mode is enabled: that's what made the difference here. Then stop the AI from reading your secret files. In Cursor, you just need to create a file named .cursorignore at the root of your project, and list in it what the agent must never open, for example:

secrets/
*.pem
*.key

And above all, never paste a password or an access key into a conversation with an AI, whatever it is. If you've already done it, consider that key burned: go back to the service that gave it to you and make a new one, it takes two minutes. I'm not going to act clever: pasting a database connection string into a session to go faster is the kind of reflex every hurried developer has had one day. Me, I use Claude Code directly in the terminal, but the risk is the same everywhere: what you type into the conversation can be kept somewhere.

A few questions remain open. Cursor does not say when the key leaked at Fireworks, or how an identifier pulled from an old 2025 log could still write to the store in 2026. And we still don't have the Fireworks version.

What strikes me is that no hacker forced a door in this whole story. He just picked up keys that others had left lying around, one by one. The safe was armored, and the code was written on a sticky note, in a 2025 notebook.

Sources

Article written with the help of Claude Code, proofread and corrected by me.

Join the conversation

You need an account to comment on this article. Creating one is free and takes under a minute.

  • The XMLTV file, free to download every day
  • Comment on articles and reply to other readers
  • Get an e-mail when an article you follow is updated

No comments yet.

Une erreur s'est produite. Cette application peut ne plus répondre jusqu'à ce qu'elle soit rechargée.Veuillez contacter l'auteur. Reload 🗙