Blog

Cursor traces its extension attack to a vendor key

Follow Cursor’s reported attack chain from read-only storage to stolen credentials and a tampered extension cache, with a Security+ paper exercise.

· 3 min read

In its update on the malicious extension incident, Cursor traces the attack back to an access key it had issued to Fireworks. The attacker used that key to read a storage bucket containing historical coding-agent traces. Some traces held credentials that had not been removed. This is a useful Security+ case because the initial permission was read-only, yet the data it exposed helped the attacker reach other systems.

Read-only describes one resource

Cursor’s investigation says the shared traces were mostly from 2025 sessions not using Privacy Mode. Fewer than 0.01% of customers had application programming interface (API) credentials exposed. Starting with those credentials, the attacker reached selected software services used by Cursor and obtained additional data. The report describes a chain of access, not a storage key magically gaining write permission.

Draw three boxes on paper: the shared storage bucket, credentials inside saved sessions, and services those credentials can access. Label each connecting arrow with the credential used at that step. The distinction to explain is between permission to read a file and the separate power of a credential written inside it. That interpretation follows the reported sequence; it is not a claim that every trace or every customer was exposed.

The tampering happened in the delivery cache

On October 29, the attacker used another stolen credential to put unauthorized copies of Anthropic’s Claude Code extension directly into the cache serving Cursor’s VS Code extension marketplace. Cursor says Anthropic did not publish the malicious copy, this was not a flaw in Anthropic’s product, and the upstream Open VSX marketplace never contained the malware. The location of the change matters: the delivery cache was a different trust boundary from the original publisher.

Cursor reports that at most 5,000 users requesting particular cached versions potentially received the malicious build during an approximately 11-hour window. “Potentially received” is not a count of confirmed infected computers. The same update also says its app and official updates were not affected and no release-signing keys were exposed. Keep those scope limits alongside the impact figures.

Map each control to the part it protects

Cursor’s published response includes disabling the vendor key, rotating exposed credentials, validating and rebuilding the marketplace store, expanding network restrictions and scanning more internal sources for secrets. Those are different actions against different links in the chain. Revoking the original storage key alone does not describe the complete response Cursor reports.

The official SY0-701 objectives list supplier and vendor attack surfaces in 2.2, malicious updates and supply-chain vulnerabilities in 2.3, incident response in 4.8 and third-party risk management in 5.3. Use the Security+ study guide for those general concepts, then assign one of Cursor’s reported response actions to each box in your diagram. This is an application exercise, not a claim that the incident will appear on the exam.

Separate vendor findings from your own assumptions

Cursor’s update says the last attacker activity it observed was November 1 and was blocked, and that it believes the incident is contained. It also warns that uninstalling a tampered extension alone may not be enough and directs affected customers to its security notice. Attribute those statements to Cursor. Your study conclusion should describe the observed route and the controls, not invent a broader compromise of Anthropic, Open VSX or every Cursor user.

Read the SY0-701 study guide

Sources