IPCRIME NEWSROOM | VERIFIED REPORTING | 15 NOVEMBER 2026
Cursor’s latest account of its security incident draws a direct line from an access key exposed through its AI infrastructure partner Fireworks to the distribution of tampered Claude Code extension builds. The company’s November 13 investigation update adds important detail to the October 29 incident and narrows what it believes was affected.
How the compromise developed
According to Cursor, an attacker used a vendor-issued access key exposed at Fireworks to copy information from a read-only storage bucket shared for model-related work. Historical coding-agent traces from sessions outside Privacy Mode were among the data. A small number contained unredacted credentials, which reportedly enabled access to selected third-party services used by Cursor. Cursor says fewer than 0.01% of customers had API credentials exposed.
The poisoned extension window
On October 29, attackers used a stolen credential to place unauthorized copies of the Anthropic Claude Code extension in the cache serving Cursor’s VS Code extension marketplace. Versions 2.1.312 and 2.1.313 were available for around 11 hours; Cursor says at most 5,000 users potentially received them. The actual number of successful host compromises has not been established publicly. Cursor stresses that Anthropic’s official extension and Open VSX were not compromised.
Scope and remediation
Cursor says its shipped application, production releases, customer code repositories and prompts from Privacy Mode sessions were not affected. It has rotated credentials, rebuilt its extension marketplace, shifted inference traffic away from Fireworks and contacted affected customers. Anyone who installed the specified versions from Cursor’s marketplace during the incident window should follow Cursor’s official remediation instructions; uninstalling an extension alone may not remove persistence established on a device.
Why this story goes beyond one extension
The incident illustrates a chain of trust failures: a partner key, a shared analytical dataset containing secrets, downstream SaaS access and a software distribution path. Defending developer infrastructure requires safeguards at each link, not only scanning the final downloaded package.
Sources and documents
- Cursor’s original incident report, November 13, 2026
- BleepingComputer investigation, November 14, 2026
IPCRIME separates confirmed statements, reported findings and unanswered questions. This article was prepared November 15, 2026; source publication and event dates may differ.