Cursor reported in an incident update published November 13, 2026 that an attacker used a Cursor-issued entry key uncovered by Fireworks AI to repeat historic coding-agent traces and credentials, then attain choose software program instruments utilized by Cursor and distribute malicious Claude Code extension builds. Lower than 0.01% of consumers had API credentials uncovered.
Fireworks’ publicity of the important thing allowed knowledge to be copied from a read-only bucket used for Cursor’s knowledge sharing with the seller. Its contents included restricted coding-agent traces, largely from 2025, from classes not utilizing Privateness Mode that Cursor was permitted to make use of for mannequin coaching. A small quantity contained unredacted credentials, offering a route into choose software-as-a-service instruments utilized by Cursor.
From these instruments, the attacker reached choose archived, non-production Cursor supply repositories, some containing enterprise contact info. Different accessed materials included restricted help knowledge and electronic mail conversations, product-usage metrics and call particulars, summaries of some coding-agent conversations, and credentials customers had despatched to Cursor’s help crew.
Affected Information and Unaffected Techniques
Amongst customers not utilizing Privateness Mode, coding-agent traces have been uncovered for fewer than 0.05%, and Cursor stated it recognized and notified every affected person immediately. The uncovered traces have been restricted to storage used for historic offline evaluation. Present buyer knowledge and reside visitors, energetic inference requests, and Privateness Mode prompts and code weren’t affected.
Cursor additionally reported no impression to the manufacturing programs and infrastructure that run its service, or to buyer code repositories. Official app updates and code shipped to purchasers and cloud companies have been unaffected: launch code sits in a separate internet hosting system the attacker by no means accessed, and no release-signing keys have been uncovered. The corporate validated all different extensions in its VS Code market and reviewed and rebuilt that retailer.
Extension Compromise and Containment
The replace connects the intrusion to Cursor’s earlier malicious Claude Code extension disclosure. On October 29, 2026, the attacker used a credential stolen from the uncovered inner knowledge to place unauthorized builds immediately on the cache serving Cursor’s VS Code extension market. Anthropic didn’t publish them, its product was not weak, and Open VSX, {the marketplace} supply, by no means carried the malware. Cursor estimates that at most 5,000 customers who requested particular cached variations doubtlessly acquired a malicious construct throughout roughly 11 hours of availability.
Cursor printed indicators of compromise and remediation materials on October 30, 2026, and started notifying affected customers and crew directors by means of in-app messages and electronic mail. Directors acquired lists of particular keys to rotate the place credentials have been uncovered, and Cursor requested the issuing third-party companies to revoke them. Prospects whose help or enterprise knowledge might have been concerned have been notified immediately.
Cursor’s monitoring programs recognized the unauthorized exercise. The corporate quickly disabled extra inner and product companies than the proof required as a precaution, disrupting sure options, and restored every after assessment. Cursor believes the incident is contained: the final attacker exercise it noticed occurred on November 1, 2026, and was blocked. Monitoring for anomalous exercise continues.
Credential Revocation and Vendor Adjustments
Cursor moved all inference visitors off Fireworks, deactivated the seller key and minimize off Fireworks’ entry to all datasets shared with it. Each uncovered Cursor credential recognized was rotated, adopted by all credentials in use throughout inner programs as a precaution, even the place there was no proof of publicity. Cursor additionally expanded community restrictions throughout service suppliers the place doable, added secret scanning to code repositories and different inner knowledge sources, redeployed manufacturing infrastructure and engaged exterior forensics specialists.
Individually, Fireworks AI’s own security bulletin studies {that a} third-party cloud supplier internet hosting its research-and-development machines skilled an incident on October 23, 2026. An unauthorized celebration reached Fireworks growth machines on October 24, 2026, and ran an unauthorized workload that led to extra customer-data impression; Fireworks detected the exercise that day. It has not detected additional profitable unauthorized entry on the shopper platform since October 25, 2026. These findings stay topic to an ongoing investigation.
Fireworks revoked and changed credentials and repair accounts, eliminated pointless elevated privileges, restricted community pathways, tightened storage and analytics entry controls, and enhanced logging and monitoring with help from Palo Alto Networks Unit 42. It disabled self-serve dataset uploads and downloads utilized by Batch Inference and dataset-based coaching throughout the investigation. Coaching SDK stays out there with zero knowledge retention, that means no coaching knowledge is saved at relaxation, whereas Lab-based coaching operates in an remoted setting.
The brand new account locations {the marketplace} compromise on the finish of a series that started with a vendor key and continued by means of credentials left in historic traces. That makes vendor-access revocation and broad credential rotation central to Cursor’s containment evaluation, whereas Fireworks’ ongoing investigation leaves the vendor-side findings open to revision.

