On Friday, 13 November 2026, Cursor confirmed a breach that should worry anyone who writes code with AI. Attackers stole an access key from Cursor's AI vendor, Fireworks AI, used it to reach inside Cursor's systems, then planted fake copies of Anthropic's official Claude Code extension in Cursor's marketplace. For about 11 hours on 29 October, up to 5,000 developers may have installed a version that tried to plant a permanent backdoor on their machines.

What Happened

It started at Fireworks AI, a company Cursor used for AI model training and inference. Cursor had given Fireworks an access key so the two companies could share data — and according to Cursor's incident report, published on 13 November, Fireworks exposed that key.

With the key, the attacker copied data from a read-only storage bucket Cursor used to share information with Fireworks. The bucket held coding agent traces — records of past AI coding sessions, mostly from 2025 — from users who had not switched on Cursor's Privacy Mode, which stops your sessions being used for training. Some traces still contained API credentials that had never been edited out: fewer than 0.05% of users had traces exposed, and fewer than 0.01% of customers had API credentials exposed — small percentages, but the credentials were live.

How One Key Opened the Door

The attacker used the stolen credentials to reach some of the software tools Cursor itself uses, and from there got hold of a limited amount of data: archived source code, customer support emails, usage figures and contact details, and summaries of some coding agent conversations.

Cursor says its monitoring spotted the activity. The company revoked the Fireworks key, rotated the exposed credentials, and temporarily switched off some internal and product services while it investigated.

Then came the second phase — the one that reached ordinary users. On 29 October, using a credential taken from the exposed data, the attacker placed malicious copies of Anthropic's Claude Code extension directly on the cache that serves Cursor's extension marketplace. The fake versions, 2.1.312 and 2.1.313, looked like a normal update of the real extension.

Extension installing on a code editor with a warning sign
The fake Claude Code extension looked like a normal update — but installing it opened a back door on the machine.

The Fake Extension: What It Did

The trojanised extension was live for about 11 hours before Cursor pulled it down. Cursor estimates at most 5,000 users who asked for those versions in that window may have received the malicious build.

Once installed, the fake extension did far more than sit quietly. According to BleepingComputer, the added code collected system information, installed an attacker-controlled SSH key, tried to switch on SSH access, and set up ways to stay on the machine after a restart — a permanent back door.

That last point matters most. Cursor has warned users directly: uninstalling the extension alone may not be enough. The back door it tried to install — reported under the name "svc-telemetry" — can survive removing the extension. Anyone who installed 2.1.312 or 2.1.313 on 29 October needs to check their machine properly, not just delete the extension.

The clean version, 2.1.314, was published on 30 October. Cursor took its marketplace offline, rebuilt it, and worked with the cloud provider hosting the attacker's server to take it down.

Anthropic's real Claude Code extension was never compromised, and neither was Open VSX, the open marketplace Cursor builds on. Cursor says there is no evidence the incident touched Privacy Mode users, customer code repositories, its production environment, the Cursor app and its updates, or any other extension in its marketplace.

It Was Not Just Cursor

The Fireworks breach reached further than Cursor. Uber has confirmed attackers stole 149,346 drivers' licence details through the same provider — told on 9 October, it confirmed on 13 October that driver data had been accessed. Uber's filings with state attorneys general put the total at 149,346 drivers, with full licence photos, names, addresses, dates of birth and licence numbers taken. TechCrunch identified the provider as Fireworks AI.

Notion also confirmed some AI chat transcripts were exposed through Fireworks — fewer than 1% of workspaces — and cut Fireworks off on 3 November. Apify and Epoch AI published their own incident reports the same day.

Security firm Wiz calls the run of October attacks "Black October". The wave hit several AI developer platforms that month — Railway, which faced about $740,000 in fraudulent charges, OpenCode, which had API keys exposed, Fireworks AI itself, and Notion. In each case, attackers went after the suppliers and tools developers trust, rather than attacking each developer one by one.

Developer checking security keys and settings on a laptop
Checking your extension version and SSH keys takes only a few minutes — and uninstalling alone is not enough.

Check Your Machine Now

If you use Cursor, run through these checks now. They take a few minutes.

  1. Check your Claude Code extension version. If it is 2.1.312 or 2.1.313 and you installed or updated it on 29 October, treat your machine as affected. Update to 2.1.314 or later straight away.

  2. Check your SSH keys. On Mac and Linux, open authorized_keys inside the .ssh folder in your home directory, and look for any key you do not recognise, including one labelled "svc-telemetry". If you find one you did not put there, remove it. Deleting the extension does not remove this key.

  3. Look for anything that starts by itself. The fake extension tried to survive a restart, so check what runs when your machine starts. Anything unfamiliar from around 29 October deserves a closer look.

  4. Rotate your keys and passwords. If Cursor notified you, or you installed one of the bad versions, change the API keys, passwords and tokens that machine could reach. A password manager makes this far less painful.

  5. Switch on Privacy Mode. Turn on Privacy Mode in the settings if you have not already. It stops your coding sessions being used for training — which is what kept those users out of the exposed data.

  6. Check what else shares the same keys. If you used the same API key or password in Cursor and in another service, change it in both places — attackers try stolen credentials everywhere.

The Bigger Lesson

Cursor says the incident is contained: no additional malicious activity has been detected since 1 November, when the last attacker activity was blocked. The company has moved all inference traffic away from Fireworks, deactivated the key it had issued to the vendor, and redeployed its production infrastructure as a precaution. It has also contacted the users it believes installed the fake versions. If Cursor has contacted you, follow its notice.

The wider lesson is a simple fact about modern software: you do not only trust the tools you install, you trust everyone those tools trust. Cursor did not lose its own keys — a key it had handed to a supplier leaked, and that one key was enough to reach its systems and then its users' machines.

For developers, the lesson is to treat extensions like software from any other source: check what you install, keep them updated, and know that a familiar name in an update is not a guarantee. For everyone else, the lesson is the same one that applies to every account you own: use different passwords, switch on every security setting a tool offers, and protect your online accounts before something goes wrong rather than after it.