Build · founder · 6 min read
Cursor's Breach Report: The Trojanized Extension Started With a Vendor's Leaked Key
Cursor traced October's poisoned Claude Code extension to a key exposed by Fireworks AI. What was taken, who is affected, and what to change.
Two weeks ago, a backdoored copy of the Claude Code extension was served to Cursor users for about 11 hours. At the time nobody could say how the attacker got it there. On November 13, Cursor published its answer, and it connects two stories we had been covering separately.
The poisoned extension was the last step in a chain that began at Fireworks AI, the same infrastructure vendor whose breach exposed Notion AI chat transcripts.
If you use Cursor, most of this report is good news for you. A small part of it should change how you work.
The chain, in plain English
Cursor’s account runs in four steps.
1. A vendor exposed a key. Cursor had issued an access key to Fireworks, which it used for model training and inference. Cursor says Fireworks exposed that key. It doesn’t say how, and Fireworks’ own November 9 bulletin doesn’t name Cursor or any other customer.
2. The key opened a storage bucket. The attacker used it to copy data from a read-only bucket Cursor used to share data with Fireworks. Inside were what Cursor calls limited coding agent traces, mostly from 2025, from sessions where Privacy Mode was off. A trace is a record of an agent session: what was asked, what the agent did.
3. Some traces contained passwords and keys. A small number had unredacted credentials in them. The attacker used those to get into some of the SaaS tools Cursor runs its business on.
4. One of those credentials reached the extension marketplace. On October 29, the attacker used it to place malicious copies of Anthropic’s Claude Code extension on the cache that serves Cursor’s marketplace. Cursor says Anthropic did not publish those copies, and that Open VSX, the upstream source, never carried the malware.
Cursor says the last attacker activity it saw was on November 1, that it was blocked, and that it believes the incident is contained.
What was taken
From the bucket: agent traces for fewer than 0.05% of users, all from sessions outside Privacy Mode. Cursor says fewer than 0.01% of customers had API credentials exposed.
From the SaaS tools the attacker got into afterwards:
- Archived, non-production Cursor source code, some of it containing business contact details
- Limited customer support data and email conversations
- Product usage metrics and contact details
- Summaries of some agent conversations
- Credentials that users had sent to Cursor support
Read that last line twice. People pasted keys into support tickets, and those keys were sitting in a help desk when the attacker arrived.
What wasn’t touched
Cursor’s list is specific, which is to its credit:
- Prompts and code from anyone using Privacy Mode
- Live traffic. Requests you send to models today were not exposed
- Customer code repositories
- Cursor’s production infrastructure, the app itself and its official updates. No release signing keys were exposed
- Other extensions and Cursor Marketplace plugins
Are you affected?
There are two separate groups, and you can be in one without the other.
The data group
Cursor says it contacted every affected customer directly. Team admins whose credentials turned up were emailed the specific keys to rotate, and Cursor asked the companies that issued those keys to revoke them. Its wording is plain: if you did not receive a notification, your data was not in the exposed traces.
So search your inbox, including spam, for mail from Cursor dated October 29 or later. If you run a team, ask whoever holds the admin seat.
The extension group
This is the “at most 5,000 users” figure. You’re in it if Cursor installed or updated the Claude Code extension between roughly 2:27 a.m. and 1:35 p.m. Pacific on October 29, on a Mac or Windows machine.
One detail has changed since the first reports. Cursor’s security notice now lists two tampered versions, 2.1.312 and 2.1.313, not one. If you checked only for 2.1.313, check again.
Cursor warns that uninstalling the extension may not be enough. Its notice includes check and cleanup scripts for both platforms. The quick manual test is to look in your SSH authorized_keys file for a key labeled “svc-telemetry”. Our earlier walkthrough covers the cleanup and what to rotate.
What to change now
Turn on Privacy Mode
This is the clearest lesson in the report. Every exposed trace came from a session with Privacy Mode off. Cursor says it was permitted to use those sessions for model training, which is how they ended up in a bucket shared with a training vendor.
Open Cursor’s settings, find Privacy Mode, and switch it on. If you pay for a team plan, ask your admin to enforce it for everyone. Whatever you lose by not contributing training data is not worth having your client’s code sitting in a third party’s storage.
Stop putting secrets where agents and support staff can read them
The credentials in this breach leaked from two places: agent sessions and support tickets. Both are avoidable.
- Never paste an API key or password into an agent chat. Put it in an environment file and tell the agent the variable’s name, not its value.
- Never send a key to a vendor’s support team. If you already have, to any vendor, rotate that key today.
- Put a spending cap on every AI provider key. Railway’s breach showed how fast stolen keys get resold.
Ask one question of every AI tool you pay for
“Which other companies receive my data, and in which settings?” Cursor users chose Cursor. Few of them knew Fireworks existed. Most vendors publish a subprocessor list. Ten minutes reading it tells you more than the marketing page will.
What Cursor changed, and what it didn’t say
Cursor says it moved all inference off Fireworks and deactivated the key, rotated every internal credential, rebuilt the extension marketplace, added secret scanning across its repos and data stores, and brought in outside forensics.
That is a solid response. The report still leaves gaps. It doesn’t mention compensation for affected users. It doesn’t explain why credentials in training traces weren’t scrubbed before the data left Cursor. And nobody has said how many of the 5,000 machines actually ran the malware.
Bottom line
For most Cursor users, nothing of yours was taken, and Cursor’s report is more candid than most. Don’t uninstall over this.
Do take the one cheap step it points to. Turn on Privacy Mode, keep secrets out of chats and tickets, and assume that anything you type into an AI tool may be stored by a company you have never heard of.
Related guides
founder · 8 min read
10 Apps Non-Technical Founders Built With Vibe Coding
Real examples of SaaS products, marketplaces, and tools built without code: what they built, which tools they used, and what it cost.
founder · 8 min read
Researchers put a virus in an agent's memory file. It spread to the next agent.
New Anthropic/EPFL research says AI agents can pass self-replicating instructions through files like CLAUDE.md. The fix is one paragraph.
founder · 9 min read
Your Agent Can Now Spend Your Money. Nobody Can Prove You Said Yes.
AWS shipped autonomous agent payments this month. The infrastructure works. The paper trail that would settle a disputed charge doesn't exist yet.
Enjoying this guide?
Get weekly practical guides, plus tool updates and implementation playbooks.