Cursor has closed its two-week investigation into the security incident that pushed a poisoned copy of Anthropic’s Claude Code extension to its users, and the trail leads to a single leaked vendor key. The company said an attacker obtained an access key it had issued to Fireworks, its AI training and inference vendor, used it to copy data from a shared storage bucket, and then used unredacted credentials in that data to plant the malicious extension. At most 5,000 users may have received the tampered build during an 11-hour window on 29 October, and Cursor says the incident is now contained.

How one key opened the door
The attack began at Fireworks, not at Cursor. An attacker obtained an access key Cursor had issued to the vendor, and the key had been exposed by Fireworks itself. With it, the attacker copied data from a read-only storage bucket the two companies used to share data for model training.
The bucket held limited coding agent traces, mostly from 2025 sessions that were not in Privacy Mode and that Cursor was permitted to use for training. A small number of those traces contained credentials that had not been redacted, and fewer than 0.01% of customers had API credentials exposed.
Starting with those credentials, the attacker reached select software tools Cursor used day to day. From them, the attacker accessed archived non-production source code, limited customer support data and email conversations, product usage metrics and contact details, plus some credentials users had sent to the support team.
The poisoned extension
On 29 October, the attacker used a stolen credential to place malicious copies of Anthropic’s Claude Code extension, versions 2.1.312 and 2.1.313, directly on the cache serving Cursor’s VS Code extension marketplace. The tampered builds carried an SSH backdoor and stayed available for about 11 hours before Cursor removed them.
At most 5,000 users who requested the specific cached versions on that day may have received the malicious build. Cursor stressed that Anthropic did not publish the copy and that the incident was not a vulnerability in Anthropic’s product, while the marketplace source, Open VSX, never hosted the malware at all.
Cursor took the marketplace offline, rebuilt it, and worked with the cloud provider hosting the attacker’s server to take it down. It also validated every other extension in the marketplace as authentic. Customers who installed the tampered versions are told to follow the steps in Cursor’s security notice, since uninstalling the extension alone may not be enough.
What stayed safe
The exposed data was limited to storage used for historical offline analysis, and live traffic was not touched. Prompts and code from users in Privacy Mode were not affected, and coding agent traces were exposed for fewer than 0.05% of users, each of whom Cursor says it has notified directly.
Production systems and infrastructure were not affected, and neither were the Cursor app, its official updates, or customer code repositories. No release signing keys were exposed, and Cursor ships its software from a separate code hosting system the attacker never reached.
What Cursor has done
Cursor has moved all inference traffic off Fireworks, deactivated the key it had issued to the vendor, and cut off its access to every dataset they shared. It rotated every exposed credential it found, highest risk first, and as a precaution rotated all credentials in use across its internal systems.
The company redeployed all production infrastructure and services, reviewed every internal service before restoring it, and added secret scanning across its code repositories and other internal data sources. It also engaged external forensics experts, expanded network access restrictions, and contacted everyone affected.
Our Read
The telling detail is how ordinary the failure was. No zero-day, no novel exploit — a vendor key, exposed by the vendor, opening a shared bucket that happened to hold credentials nobody had redacted. The whole chain, from a training-data share to a poisoned extension in 5,000 users’ editors, ran on access that should have been scoped, rotated and scanned.
That is the lesson the wider industry keeps relearning: the riskiest credential is often not a company’s own, but one it handed to a partner. Cursor’s response — cutting Fireworks off entirely, rotating everything, rebuilding the marketplace — is the right scale of answer, but it comes after the fact.
The episode also closes a loop this site has followed for a month. Fireworks’ breach disclosure widened on 9 November to cover datasets and model weights, and Notion then confirmed its AI chats had leaked through the same vendor. Cursor’s findings tie the marketplace attack to the same root cause — one vendor’s exposed key, reaching three companies.
Why It Matters
For anyone who installed or updated the Claude Code extension from Cursor’s marketplace on 29 October, the message is direct: check whether you received versions 2.1.312 or 2.1.313, and follow Cursor’s remediation steps. The company warns that simply uninstalling the extension may not remove everything the backdoor touched.
For the industry, the incident is a reminder that extension marketplaces are trust infrastructure. Developers install what a marketplace serves without inspecting it, so a single compromised cache entry reaches thousands of machines in hours. Vendors that share data with AI partners need to treat those shared credentials as production secrets, not training conveniences.
Background
Cursor first disclosed the incident on 29 October, the same day the tampered builds appeared, and published a security notice with indicators of compromise the next day. Fireworks disclosed its own breach on 9 November, and Notion confirmed its AI chats had leaked through the vendor on 10 November. Cursor says it will update its post if anything important changes.
Source: Cursor’s security update


0 Comments