Cursor security incident: how a vendor key led to malware
The Cursor security incident began with a leaked Fireworks key and ended in a backdoored Claude Code extension. What happened and how to check your machine.
Cursor published the full write-up of its security incident on November 13, 2026, and it is one of the clearest supply chain stories I've read this year. The Cursor security incident did not start with a bug in Cursor. It started with an access key Cursor had issued to Fireworks AI, a model inference vendor. That key leaked, and two weeks of credential hopping later, a tampered copy of Anthropic's Claude Code extension was sitting in Cursor's extension marketplace cache with an SSH backdoor inside.
I do security work at erxes, and we hand keys and data to vendors all the time, like everyone else. So I read this less as a Cursor story and more as a map of how a vendor breach becomes your breach. Here is what happened, who needs to act, and what I'd change in my own setup.
How the Cursor security incident unfolded
According to Cursor's incident update, the chain went like this:
- An attacker got an access key Cursor had issued to Fireworks. Cursor says the key was exposed by Fireworks. The attacker used it to copy data from a read-only storage bucket Cursor used to share data with the vendor.
- That bucket held coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which Cursor was allowed to use for model training. A small number of those traces contained credentials that had never been redacted. Cursor puts it at less than 0.01% of customers with exposed API credentials.
- With those credentials the attacker reached some of the SaaS tools Cursor uses. From there they got into archived non-production source repositories, limited support data and emails, usage metrics and contact details, summaries of some agent conversations, and credentials users had sent to Cursor support.
- On October 29, using a stolen internal credential, the attacker placed malicious copies of the Claude Code extension directly on the cache that serves the Cursor VS Code extension marketplace. The upstream source, Open VSX, never had the malware. Anthropic did not publish it, and Cursor says it was not a vulnerability in Anthropic's product.
- At most 5,000 users who requested those cached versions on October 29 may have received the malicious build. It was live for about 11 hours. The last attacker activity Cursor saw was on November 1, and it was blocked.
The part that stands out to me is step 2. Nobody had to break crypto or find a zero-day. Developers pasted secrets into agent sessions, the sessions became training data, the training data went to a vendor, and the vendor lost a key.
What the malicious Claude Code extension did
Cursor's October 29 security notice lists the payload in versions 2.1.312 and 2.1.313 for Windows and macOS. It:
- sent host details to the attacker: hostname, username, OS, architecture, home directory, environment variable names, working directories, Git email and public IP
- added an attacker SSH key labeled
svc-telemetrytoauthorized_keys, tried to turn on the local SSH service and open a reverse SSH tunnel - installed persistence with a LaunchAgent on macOS, or a scheduled task with a Startup-script fallback on Windows, retrying the tunnel about every 60 seconds
Turning on SSH needs admin rights, so on most machines that step failed and the tunnel never worked. The attacker's server has also been taken offline. But the key and the persistence scripts still got written, and the notice is blunt that uninstalling the extension may not remove them.
How to check your machine
On macOS, Cursor's notice gives this one-liner:
grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted"
The notice also has a PowerShell check for Windows, remediation scripts for both platforms, and file hashes and network indicators for your EDR. If the key is there, isolate the machine, run the script, consider a rebuild, and rotate every credential and session that machine could reach, from a clean system.
The Fireworks breach is bigger than Cursor
Cursor was not the only victim. TechCrunch reported on November 13 that Uber is notifying drivers whose personal and driver's license data was stolen through the same Fireworks breach. State filings show 2,567 people in Massachusetts, 3,430 in South Carolina and 1,873 in Rhode Island, with more states pending. Notion has also confirmed its data was taken. Fireworks blamed a financially motivated group that sells stolen tokens for LLM access.
That motive matters. If the goal is reselling model access, then every API key sitting in a log, a trace or a training set has a resale price. Agent traces are full of those.
What I'd change after reading this
Cursor's fix list is solid: it moved all inference off Fireworks, rotated every credential it found and then all internal credentials as a precaution, restricted where vendor credentials can be used from, added secret scanning across repos and other internal data, rebuilt the extension marketplace and redeployed production. Here is what I'm taking from it for my own team and for the agent payment risks I wrote about yesterday:
- Treat agent traces as secret-bearing data. Run secret scanning on logs and traces before they leave your boundary, not only on git repos.
- Bind vendor keys to a network location. A key that only works from the vendor's IP range is far less useful once it leaks.
- Share the minimum, for the shortest time. A read-only bucket still leaked a lot. Expire shared datasets.
- Don't trust a cache more than its source. The marketplace source was clean, the cache was not. If you mirror packages or extensions, verify hashes against upstream.
- Stop sending credentials to support. Users did it, and those credentials were exposed too. Give people a safer channel and redact on intake.
None of this is new advice. The incident just shows what it costs when one link skips it. I said something similar about self-hosted n8n patching: the dangerous gaps are usually in the boring plumbing between systems.
FAQ
Was I affected by the Cursor security incident?
Only if you installed or updated Anthropic.claude-code 2.1.312 or 2.1.313 from the Cursor marketplace on macOS or Windows between 2:27am and 1:35pm PT on October 29, 2026, or if Cursor notified you directly about exposed traces, credentials or support data. Cursor says it contacted every affected customer.
Was Privacy Mode data exposed?
Cursor says prompts and code from Privacy Mode sessions were not affected. The exposed traces came from non-Privacy-Mode sessions, and fewer than 0.05% of users had traces exposed.
Is the Cursor app itself compromised?
Cursor says no. It ships the app from a separate code hosting system the attacker never accessed, no release signing keys were exposed, and every other extension in the marketplace was validated as authentic.
Sources
- An update on our recent security incident, Cursor (November 13, 2026)
- Security notice for malicious Anthropic.claude-code 2.1.312 and 2.1.313, Cursor (October 29, 2026)
- Uber confirms hackers stole drivers' license information, TechCrunch
- Cursor says stolen Fireworks AI key led to trojanized Claude Code extension, The Hacker News
Cover image: Cursor
Related articles
Microsoft agent-framework PyPI compromise: what to check
The agent-framework PyPI compromise put a credential stealer in versions 1.26.1 and 1.26.2. How it runs, how to check your machines and what to rotate.
WPPConnect npm compromise: what to check and rotate
The WPPConnect npm compromise shipped a secret stealer through trusted publishing. How the attack worked, how to check your systems and what to rotate.
FBI ARTEX alert: how to defend against AI-driven attacks
The FBI ARTEX alert warns AI agents now run intrusions in under three hours. What ARTEX does, known campaigns, and the defenses to put in place first.
Securing AI agents: lessons from Uber's Secure AI framework
Securing AI agents takes more than API keys. How Uber's Secure AI framework checks if an agent's action fits its purpose, and what small teams can copy.
n8n security update: 5 vulnerabilities to patch now
The n8n security update of November 13 fixes two high-severity account takeovers and three more flaws. Versions to install and what to check on self-hosted n8n.
ChatGPT Wallet security: how AI agent payments work
ChatGPT Wallet lets AI agents pay with one-time tokens and a biometric check per purchase. How it works, where the risk sits, and what merchants must change.