Cursor finally connected the dots on Friday, November 13. The AI code editor published a full incident report explaining how a booby-trapped copy of Anthropic’s Claude Code extension ended up in its marketplace on October 29. The trail starts at Fireworks AI, a vendor Cursor used to train and run models.

“An attacker obtained an access key we had issued to Fireworks. The key was exposed by Fireworks,” Cursor wrote in its November 13 update. That one sentence ties together two incidents that looked separate for two weeks: the Fireworks breach Cursor’s users had read about, and the fake extension that showed up in Cursor’s own store.

Up to 5,000 users may have installed the bad build. Cursor says the malicious versions sat in its marketplace cache for about 11 hours, from 2:27 a.m. to 1:35 p.m. Pacific time on October 29, and warns that simply uninstalling the extension may not clean up the damage.

How one leaked key became a marketplace attack

The chain ran through shared storage. The exposed key gave read access to a storage bucket Cursor used to share data with Fireworks. The attacker copied historical coding-agent traces out of it, mostly from 2025, from sessions that were not in Privacy Mode and that Cursor was allowed to use for model training.

A small number of those traces contained credentials that had not been redacted. Cursor puts the numbers low: traces were exposed for fewer than 0.05% of users, and API credentials for fewer than 0.01% of customers. Cursor says it identified and notified each affected user directly.

Those credentials were the pivot. Starting from them, the attacker reached some of Cursor’s software-as-a-service tools, then archived, non-production source repositories, some holding business contact information. From there, the attacker placed modified copies of the Anthropic.claude-code extension, versions 2.1.312 and 2.1.313, directly on the cache that serves Cursor’s VS Code extension marketplace. The fake builds never appeared on Open VSX, the upstream registry, and Microsoft says they were not distributed through its Visual Studio Marketplace or the @anthropic-ai/claude-code npm package.

What the fake extension did

The tampered build ran a small piece of malicious code the moment the extension loaded, with no click needed from the user. Researchers cited by BleepingComputer say it gathered the username, computer name, public IP address, and Git email, then tried to add the attacker’s own SSH key to the machine, switch on remote login, and keep a tunnel open back to the attacker.

Cursor published indicators of compromise and remediation material on October 30, a day after the bad builds went up, and began notifying affected users and team administrators through in-app messages and email. Its current advice is blunt: update to version 2.1.314 and follow the removal steps and scripts in its security notice. There is also no confirmed count of successful workstation backdoors, only the estimate of up to 5,000 potential recipients, and neither Cursor nor Anthropic has explained how the attacker was able to publish the build under Anthropic’s name.

What Cursor says was not affected

Cursor’s report draws a firm line around what it says stayed safe. Prompts and code in Privacy Mode were not exposed. Customer code repositories were not affected. Production systems, the Cursor app itself, official updates, and release signing keys were not touched. Live traffic, active inference requests, and other marketplace extensions were not affected either.

The company says its monitoring systems spotted the unauthorized activity, and it temporarily disabled more internal and product services than the evidence required as a precaution, disrupting some features, then restored each after review. Cursor believes the incident is contained: the last attacker activity it observed was on November 1, and it was blocked.

There is one attribution caveat worth stating plainly. That the key was “exposed by Fireworks” is Cursor’s account. Fireworks’ own security bulletins, last updated November 9, do not mention Cursor. Fireworks’ November 9 update describes a separate path in its own October investigation: a third-party cloud provider hosting its research machines had an incident on October 23, which led to wider impact on account and billing information and, for a limited set of accounts, secrets, datasets, and incomplete model weights. Fireworks brought in Palo Alto Networks Unit 42 for forensics and says its investigation is ongoing.

What Cursor changed after the breach

Cursor says it moved all inference traffic off Fireworks, deactivated the vendor key, cut off Fireworks’ access to shared datasets, rotated the exposed credentials and then all internal credentials, rebuilt its extension marketplace, expanded secret scanning, and redeployed production infrastructure as a precaution.

Why this matters

The Cursor breach is a supply-chain story wearing a malware costume. The fake extension was the visible damage, but the real failure happened earlier, in a vendor relationship: a key issued to a partner, a shared bucket of agent traces, and credentials sitting unredacted inside training data.

Agent traces deserve a second look here. They are the working notes of AI coding sessions, and they can hold more than code. Cursor’s own report shows that a trace can carry a live credential, and that a credential can carry an attacker from a read-only bucket into a company’s tools and then into its distribution channel. The lesson isn’t that traces should never be kept. It’s that anything kept for training needs the same redaction and access control as production data, because an attacker will treat it that way.

For developers, the practical takeaway is narrow and urgent. If you installed Claude Code versions 2.1.312 or 2.1.313 from Cursor’s marketplace on October 29, assume the machine needs the full cleanup in Cursor’s notice, not just an uninstall.

FAQ

What happened in the Cursor security breach?

Cursor says an attacker used an access key exposed at Fireworks AI to copy coding-agent traces from a shared storage bucket, then used stolen credentials to plant a fake Claude Code extension in Cursor’s marketplace on October 29, 2026.

How many Cursor users got the fake Claude Code extension?

Cursor estimates at most 5,000 users who requested the affected cached versions received the malicious builds, versions 2.1.312 and 2.1.313, during about 11 hours on October 29.

What did the fake Claude Code extension do?

Researchers cited by BleepingComputer say it ran code as soon as the extension loaded, collected the username, computer name, public IP address and Git email, then tried to add the attacker’s SSH key and open a remote-access tunnel.

What should Cursor users do after the breach?

Cursor says users should update to version 2.1.314 and follow the removal steps and scripts in its security notice. Users who installed versions 2.1.312 or 2.1.313 on October 29 should treat the machine as compromised until cleaned.

Sources: BleepingComputer, Unite.AI, Cursor