
This post is a translation of the Spanish original.
The extension was legitimate. What was compromised was the cache
The first thing I did when I read Cursor's update was look for a name in a file: svc-telemetry, in ~/.ssh/authorized_keys. I don't use Cursor daily, but I wanted to see what the advisory asked for, and that's what it asks. And it seemed to me the most revealing part of the whole story.
What happened
Cursor had issued an access key to Fireworks AI, the provider handling part of its inference. That key was exposed on Fireworks' side and the attacker used it to copy data from a read-only bucket both companies shared. Inside there were unredacted credentials. With one of them, on 29 October the attacker placed tampered copies of the Claude Code extension, versions 2.1.312 and 2.1.313 for macOS and Windows, in the cache that serves Cursor's extension marketplace.
They were available for about 11 hours. Cursor estimates that at most 5,000 users who requested those specific versions could have received the malicious build. Anthropic did not publish those copies and its official distribution was not affected.
The payload ran when the extension loaded, without clicking anything. It collected machine data, tried to add an attacker SSH key, enable the SSH service, open a reverse tunnel and leave persistence with a LaunchAgent on macOS or a scheduled task on Windows. According to Cursor, because of bugs in the code itself, enabling SSH failed unless Cursor had been launched as administrator. But the key and the persistence could stay. That's why they insist that uninstalling the extension isn't enough.
My reading: the weak link wasn't the code
What strikes me is the chain. The original extension was fine. The source repository, according to Cursor, was too. What failed was an intermediate piece almost nobody thinks about: a cache in front of the marketplace, protected by credentials that ended up in a bucket shared with a third party.
From the systems side this is very recognisable. We set up a mirror, a package proxy or a cache to gain speed, and from then on we trust what it serves without checking again. If signature or hash verification is done against the origin and not against what actually reaches the machine, the cache becomes the ideal place for a swap.
And then there's the type of tool. A coding agent's extension lives on the developer's laptop, with access to repositories, keys and open sessions. It's almost the worst possible place for something to update itself.
My take: IDE extensions with agents should be treated as production packages. That means, at minimum:
- Pinned versions and updates with a few days' delay, not the minute they're published.
- Integrity verification against what gets installed, not against what's supposed to be at the origin.
- Vendor credentials with minimal scope and nothing unredacted in shared buckets.
I'm clear on the best counter-argument: if you delay updates, you also delay security patches, and some extensions fix serious flaws. True. But in this particular case, a 24-hour delay would have kept every affected user out, because the window was 11 hours. For most teams, I prefer that risk to the opposite.
What I'd do if you use Cursor
If you installed or updated the Claude Code extension in Cursor on 29 October, I'd follow the security notice to the letter: look for the svc-telemetry key, check persistence and tunnels, use their cleanup scripts and rotate credentials and sessions from another trusted machine. Having those versions installed doesn't by itself mean being compromised, but I wouldn't take it as fine without checking.
What keeps going round my head is something else. We install more and more tools that bring their own marketplace, their own cache and their own provider behind them. How many of them could you explain where they really get what they install from?


