Threat

Stolen Fireworks AI Key Led to Cursor Data Breach and Trojanized Claude Code Extension

Page published · Page updated

Dossier

Earliest dated coverage: 15 Nov 2026 · First observed: 15 Nov 2026 · Latest dated coverage: 15 Nov 2026

Coverage timeline

15 Nov 2026thehackernews.com

Single-source incident — one report is available.

Why it matters

The Cursor compromise highlights how exposed partner credentials and compromised AI-extension distribution can endanger shared data and developers who trust marketplace updates.

Cursor says an attacker used an access key exposed by Fireworks AI to copy data from a read-only storage bucket shared with the provider. The reported breach was linked to the distribution of a trojanized official Claude Code extension as a legitimate update through Cursor Marketplace.

campaign

Summary

Cursor attributed a data breach and subsequent Claude Code extension compromise to an access key exposed through Fireworks AI. According to the account, the attacker used the key to copy a shared read-only storage bucket, recovered credentials from its contents, accessed selected SaaS tools used by Cursor, and ultimately obtained a credential that enabled tampering with the cache serving Cursor Marketplace. The reported access path is credential-driven; no specific software vulnerability is identified.[0]

The initial bucket contained limited coding-agent traces from sessions that were not in Privacy Mode and were permitted for model training, with most traces dating to 2025. Unredacted secrets exposed API credentials belonging to less than 0.01% of customers. Subsequent access reached selected archived non-production Cursor source repositories, some containing business contact information; limited support data and email conversations; product usage metrics and contact details; summaries of coding-agent conversations; and credentials users had submitted to support.[0]

On October 29, 2026, the attacker placed trojanized Claude Code extension builds directly in Cursor's marketplace-serving cache. The builds attempted to install a persistent SSH backdoor. Cursor estimated that at most 5,000 users potentially received malicious builds during approximately 11 hours of availability and notified affected users who installed versions 2.1.312 or 2.1.313 during that window. Potential delivery, installation, and successful backdoor establishment should not be treated as equivalent outcomes.[0]

Cursor reported removing the builds, taking the marketplace offline, coordinating takedown of the attacker's hosted server, revoking and rotating credentials, and investigating deeper access. It also deactivated the Fireworks key, moved inference traffic away from the provider, and terminated access to shared datasets. Cursor said it had detected no additional malicious activity since November 1, 2026, and had no evidence of impact to Privacy Mode users, customer code repositories, its production environment, its applications or official updates, or other marketplace extensions and plugins. These are vendor-reported findings, not independently corroborated conclusions in the supplied evidence.[0]

The article places the Cursor intrusion within a broader Fireworks-related data-exposure context, also reporting theft affecting Uber and Notion. Uber's quoted disclosure identifies only an unnamed service provider; the article attributes identification of that provider as Fireworks to TechCrunch. The supplied account does not establish that the same attacker conducted all of these incidents.[0]

Attack chain

  1. 1. Obtain the provider-issued access key: The attacker obtained an access key that Cursor had issued to Fireworks AI and that Cursor said Fireworks exposed. The report does not describe the technical method used to compromise Fireworks or acquire the key.[0]
  2. 2. Copy shared coding-agent traces: Using the key, the attacker copied data from a read-only storage bucket used to share information with Fireworks. The bucket contained limited, training-permitted traces from non-Privacy Mode sessions, including credentials that had not been redacted.[0]
  3. 3. Reuse exposed credentials for further access: Cursor said the attacker used credentials from the exposed material as a starting point to reach selected SaaS tools used by the company. This connected the initial shared-data exposure to further access beyond the storage bucket.[0]
  4. 4. Access additional organizational data: The subsequent access reached selected archived non-production source repositories, limited support records and email conversations, usage metrics and contact details, coding-agent conversation summaries, and credentials supplied to Cursor support.[0]
  5. 5. Place malicious builds in the marketplace cache: On October 29, 2026, the attacker used a stolen credential from exposed internal-system data to place trojanized Claude Code extension builds directly in the cache serving Cursor Marketplace. The report describes cache tampering, rather than a compromise of Anthropic's own release process.[0]
  6. 6. Deliver the extension and attempt persistence: Users requesting the affected cached versions could receive the trojanized extension as a legitimate update. The malicious builds attempted to install a persistent SSH backdoor. Cursor bounded potential delivery at 5,000 users during approximately 11 hours; the account does not provide a confirmed successful-backdoor count.[0]

Disclosure timeline

DateEvent
2026-10-29The attacker placed trojanized Claude Code extension builds in Cursor's marketplace-serving cache. Cursor said the builds were available for approximately 11 hours before removal and that at most 5,000 users potentially received them.[0]
2026-11-01Cursor subsequently identified this as the date since which it had detected no additional malicious activity. The account does not identify this date as the completion of every response action.[0]
2026-11-15The Hacker News published the supplied account, citing a Friday Cursor update that connected the Fireworks-exposed key, subsequent data access, and October 29 extension-cache compromise.[0]

Affected versions and patch status

ProductAffectedPatch status
Anthropic Claude Code extension distributed through Cursor MarketplaceTrojanized cached builds associated with versions 2.1.312 and 2.1.313 installed from Cursor Marketplace during the approximately 11-hour exposure window on October 29, 2026. Cursor estimated at most 5,000 potential recipients. The account does not establish that every copy of either version, or copies distributed through other channels, was malicious.Cursor removed the malicious builds and took its extension marketplace offline. A fixed extension version or specific patch release is not identified in the supplied account; the reported containment was removal and distribution shutdown.[0]

Key takeaways

  • Read-only access did not prevent downstream compromise: the shared bucket contained reusable credentials that the attacker used to reach additional services and data.[0]
  • The reported software-distribution compromise occurred in Cursor's marketplace-serving cache. It should not be generalized into a confirmed compromise of Anthropic's release infrastructure, Cursor's application updates, or all marketplace content.[0]
  • The less-than-0.01% figure concerns customers whose API credentials appeared in exposed traces; the 5,000-user upper bound concerns potential recipients of malicious extension builds. Neither figure is a confirmed count of successfully backdoored endpoints.[0]
  • Cursor's finding of no evidence of impact to Privacy Mode users and other listed environments is an investigation-status statement. It should remain distinct from the affirmative findings of shared-data theft and malicious extension distribution.[0]
  • The supplied evidence consists of one news account relaying vendor statements and other reporting. It names no threat actor and supplies no actionable attacker domains, addresses, hashes, or backdoor paths, so attribution and indicator-based hunting remain limited.[0]

Defensive actions

  • Scope potentially affected developer endpoints by extension version, distribution channel, and installation date.: Prioritize Claude Code versions 2.1.312 and 2.1.313 installed through Cursor Marketplace during the October 29 exposure window, and reconcile findings with Cursor's affected-user notifications. The incident was tied to particular cached builds, so version numbers alone do not establish malicious installation.[0]
  • Investigate potentially affected endpoints for the reported persistent SSH backdoor.: The trojanized extension attempted to establish SSH persistence. Endpoint investigation is needed to distinguish potential receipt of a malicious build from successful backdoor installation; the supplied report provides neither a success count nor specific backdoor artifacts.[0]
  • Revoke and rotate credentials identified in exposed traces, affected support submissions, or incident notifications.: Unredacted credentials in shared traces enabled the initial expansion of access, and additional credentials were present in support-related material. Cursor reported revocation and rotation and assisted affected users with credential revocation.[0]
  • Extend the access investigation beyond the shared bucket to the selected SaaS services and data collections described in the incident.: The attacker reused exposed credentials to access SaaS tools and additional organizational data. Cursor reported hunting for deeper access and identifying further access, making a bucket-only review insufficient for the documented intrusion path.[0]
  • For the marketplace operator, remove tampered cached builds, interrupt their distribution, and coordinate takedown of attacker-hosted infrastructure.: Cursor reported removing the malicious extension builds, taking the marketplace offline, and working with the hosting cloud provider to take down the attacker's server. These actions addressed the delivery channel and supporting infrastructure described in the account.[0]
  • For the affected service operator, terminate the compromised provider key and its access to shared datasets.: Cursor reported deactivating the Fireworks-issued key, cutting off dataset access, and moving inference traffic off Fireworks. These steps addressed the provider-linked access path used at the start of the intrusion.[0]