Inter-office memo

To
Everyone using Cursor
From
Mo, Chief Executive
Date
November 14, 2026, 3:09 AM CST
Re
Cursor’s security incident

Cursor’s breach began with a key Fireworks AI exposed. Who’s affected, and what to do

On November 13, 2026, Cursor said its October breach began with an access key it had issued to Fireworks AI, which Fireworks exposed: the attacker copied agent traces from sessions not in Privacy Mode and used credentials in them to reach support data, archived code and the extension marketplace. If Cursor hasn’t contacted you, it says your traces weren’t taken; Privacy Mode code and app updates weren’t touched.

What Cursor said on November 13

On November 13, 2026, Cursor published an account of the incident it first disclosed on October 29, when tampered copies of Anthropic’s Claude Code extension turned up in its extension marketplace. After two weeks of investigation, it says the extension was the last link in a longer chain, and that “This incident involved data we stored and shared, credentials we issued, and systems we operate.”

“The vast majority of customers were not affected,” Cursor says, “and we have contacted each affected customer directly with the details that apply to them.” It believes the incident is contained: “The last attacker activity we have seen was on November 1, and it was blocked.”

How it happened

Cursor’s account, in order:

  1. It started at Fireworks. “An attacker obtained an access key we had issued to Fireworks,” the training and inference vendor Cursor used. Cursor says “The key was exposed by Fireworks,” and the attacker used it “to copy data from a read-only storage bucket we used to share data with them.”
  2. The bucket held agent traces. They were “limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which we were permitted to use for model training.” A small number contained credentials “that had not been redacted.”
  3. Those credentials opened Cursor’s tools. “Starting with those credentials, the attacker reached select SaaS tools used by Cursor,” and through them archived code, support data and more, as the next section lists.
  4. Another credential reached the marketplace. On October 29, “using a stolen credential from the exposed data from our internal system,” the attacker put tampered copies of the Claude Code extension “directly on the cache used to serve the Cursor VS Code extension marketplace,” for about 11 hours. “Anthropic did not publish this copy,” Cursor says, and “The marketplace source, Open VSX, never had the malware.”

The dates, from Cursor’s and Fireworks’ pages:

Date What happened
Oct 6 Fireworks finds its internal credentials being used without permission
Oct 23–24 A cloud provider hosting Fireworks’ research machines is breached, and the intruder reaches Fireworks’ development machines
Oct 29 Tampered Claude Code extensions 2.1.312 and 2.1.313 are in Cursor’s marketplace, from about 2:27 a.m. to 1:35 p.m. Pacific
Oct 29–30 Cursor posts its security notice and starts notifying users and team admins in the app and by email
Nov 1 The last attacker activity Cursor saw, which it blocked
Nov 9 Fireworks updates its bulletin, and Notion tells customers that its chats were exposed at Fireworks
Nov 13 Cursor publishes its account

Cursor doesn’t say when its key was exposed or which of Fireworks’ incidents it was part of. Fireworks’ bulletins, last updated November 9, don’t mention Cursor’s key. Its October bulletin did say that “environment variables, including API keys and tokens stored as secrets, were captured in certain affected accounts.”

What was exposed, and what wasn’t

Data or system Affected? Cursor’s words
Agent traces, from sessions not in Privacy Mode Yes, for “fewer than 0.05%” of users “mostly from 2025”
API credentials in those traces Yes, for “less than 0.01% of customers” “had not been redacted”
Credentials users sent to Cursor’s support team Some “some credentials users had sent to the Cursor support team”
Support data and email Some “limited customer support data and email conversations”
Cursor’s own code Some “select archived non-production Cursor source code repositories (some containing business contact information)”
Usage data Some “product usage metrics and contact details”
Conversation summaries Some “summaries of some coding agent conversations”
The Claude Code extension in Cursor’s marketplace Tampered, October 29 “At most 5,000 users” may have received it
Prompts and code in Privacy Mode No “Your prompts and code if you use Privacy Mode”
What you send Cursor now No Including “Active inference requests”
Cursor’s production systems No “The production systems and infrastructure that run Cursor”
The Cursor app and its updates No “no release signing keys were exposed”
Your code repositories No “Customer code repositories”
Other extensions and plugins No “We validated that all other extensions were authentic”
Cloud Agents and Grok Bot No Along with “all other SpaceX SI products”

The app line matters most for whether to keep using Cursor: “We ship Cursor from a separate code hosting system that the attacker never accessed.”

Who’s affected, and what to do

If you What Cursor says What to do
Use Privacy Mode Your prompts and code weren’t affected Nothing about traces
Don’t use Privacy Mode, and got no notice “your data was not in these traces” Consider turning Privacy Mode on
Got a notice that your traces were exposed Fewer than 0.05% of users Replace any key or password those sessions held
Are a team admin and got a list of keys Cursor has asked the issuers to revoke them Rotate every key on the list
Sent a key or password to Cursor support Some were taken; those customers were notified Replace it, notice or not
Installed the Claude Code extension from Cursor’s marketplace on October 29 At most 5,000 users may have got the tampered build Check for the backdoor’s SSH key

Step by step:

  1. Look for Cursor’s notice. It went to affected users and team admins “through in-app notices and email.” Team admins whose keys were exposed got “a list of the specific keys to rotate.”
  2. Replace the listed keys where they’re issued. Cursor “asked the third-party services that issued those credentials to revoke them,” so a listed key may already have stopped working. Make a new one with the service that issued it and update whatever used the old one.
  3. Replace anything you sent support. That’s our advice, not Cursor’s: Cursor says it notified the customers whose support credentials were taken, but a key that sat in a support thread is safer replaced.
  4. Check the extension, if you used Claude Code in Cursor on October 29. On a Mac or Linux, run grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted". Cursor warns that “Uninstalling the extension alone may not be enough.” Our extension memo has the Windows check and the cleanup scripts.
  5. Turn on Privacy Mode. Open Cursor Settings, click General, and turn on Privacy Mode, per Cursor’s help. “For teams, Privacy Mode is enabled by default,” and admins can enforce it at cursor.com/dashboard.
  6. Be wary of mail about the incident. Contact details and support email were among what was taken, which is the material targeted phishing is made from.

Privacy Mode, and why it mattered here

Cursor’s data use page sets out the difference. With Privacy Mode on, “Customer Data will not be used for training by Cursor,” and Cursor keeps zero data retention agreements with its model providers. With it off, “we may use and store codebase data, prompts, editor actions, code snippets, and other code data and actions to improve our AI features and train our models.” The bucket the attacker copied held the second kind, shared with a vendor, with some credentials left in.

The lesson reaches past Cursor. Data kept for training is data kept, sometimes with a vendor’s vendor; Notion’s chats reached Fireworks through a feature that stored them too. Other agents have the same switch: OpenAI’s Improve the model for everyone, which covers Codex, and Google’s Keep Activity for Gemini. And a key typed into any chat is a key someone else may store, so keep them out, as our help guide describes.

What Cursor has changed

Cursor lists these steps:

  • “Moved all inference traffic off Fireworks, deactivated the key we had issued to it, and cut off its access to all data sets we shared with it.”
  • Rotated every exposed credential, “and as a precaution rotated all credentials in use across our internal systems, even where we had no evidence of exposure.”
  • Validated every other extension and “reviewed and rebuilt” its extension marketplace.
  • Restricted where its service providers’ credentials work, where possible, “so compromised credentials can’t be used outside their intended location.”
  • Added secret scanning to its code repositories and other internal data.
  • Redeployed all production infrastructure, reviewed every internal service before restoring it, and engaged outside forensics experts.

It also says that during the response it “temporarily disabled internal and product services well beyond what the evidence required, which temporarily disrupted certain Cursor features,” restoring each “once it was safe to do so.”

What hasn’t been said

  • How and when Fireworks exposed the key, and which of its incidents it was. Fireworks hasn’t commented on Cursor’s account on its own pages.
  • How many people the percentages mean. Cursor gives shares, not counts.
  • Whether any exposed key was used before it was revoked.
  • Whose conversations the summaries came from, and whether any were Privacy Mode sessions. Cursor’s list of what was taken doesn’t say.
  • Which features ran on Fireworks, and what moving off it changed. Cursor doesn’t name the features it disrupted.
  • What its forensics experts found. Cursor says it will update its post “If anything important changes.”

If you’re deciding whether to stay

Cursor’s case for staying is that the app, its updates and your repositories weren’t affected, and that it has rotated everything and moved off the vendor. If you stay, turn on Privacy Mode. If you’re weighing a move, our Cursor alternatives guide compares the options by price and by what they run on.

Enjoy is one of them. It has no AI model of its own and never sees your conversations: it runs each lab’s own agent on your computer, on the plan you already have, and the agent sends its work to the provider you signed in with, under that provider’s rules (what AI providers see). That makes each provider’s training setting the one to check. To send nothing at all, run a local model, which Enjoy runs through OpenCode.

9.Questions.

Was Cursor hacked?

Yes, by its own account. On November 13, 2026, Cursor said an attacker used an access key it had issued to Fireworks AI, which Fireworks exposed, to copy agent traces, then used credentials in them to reach some of Cursor’s tools, archived code and support data, and its extension marketplace. Cursor says its production systems, app updates and customers’ code repositories weren’t affected.

Was my Cursor data exposed?

Only if Cursor told you so. It says agent traces were exposed for “fewer than 0.05%” of users, all from sessions not in Privacy Mode, and that “If you did not receive a notification, your data was not in these traces.” With Privacy Mode on, Cursor says your prompts and code weren’t affected.

Do I need to rotate my API keys after the Cursor breach?

Only the ones Cursor listed, by its account. Fewer than 0.01% of customers had API credentials exposed, and Cursor emailed their team admins a list of the keys to rotate and asked the services that issued them to revoke them. If you ever sent a key or password to Cursor’s support team, replace it anyway; some such credentials were taken.

How did the malicious Claude Code extension get into Cursor’s marketplace?

With a credential taken from the same exposed data. On October 29, the attacker put tampered copies of Anthropic’s Claude Code extension “directly on the cache used to serve the Cursor VS Code extension marketplace,” Cursor says. Anthropic didn’t publish them, Open VSX never had the malware, and at most 5,000 users may have got them. How to check.

Does Privacy Mode protect my code in Cursor?

In this incident, yes. The bucket the attacker copied held traces from sessions not in Privacy Mode, which Cursor was “permitted to use for model training,” and Cursor says Privacy Mode prompts and code weren’t affected. Turn it on in Cursor Settings, under General; it’s on by default for teams.

Is Cursor safe to use now?

Cursor believes the incident is contained. It says the last attacker activity was on November 1 and was blocked, that it has rotated every internal credential and redeployed production, and that its app ships from a separate system the attacker never reached. It hasn’t published its outside forensics team’s findings.

Enjoy Agentic Industries · Checked against its sources on