The AI-powered code editor Cursor has confirmed suffering a security breach that enabled threat actors to publish malicious extensions through its marketplace after obtaining credentials from Fireworks AI, a third-party AI infrastructure vendor. The incident also granted attackers access to internal systems and customer information, including unredacted API keys embedded in historical AI coding-agent traces that were subsequently leveraged to compromise additional infrastructure.

Cursor released a detailed incident report on Friday outlining how the attackers distributed compromised versions of Anthropic's Claude Code extension via Cursor's marketplace on October 29. The malicious builds exposed as many as 5,000 developers to malware designed to establish persistent SSH backdoors on their systems.

image

Cursor is a Visual Studio Code-based AI-assisted code editor that enables software developers to write, modify, and debug applications with the help of AI tools. The platform is built on Microsoft's open-source VS Code foundation.

How the Breach Unfolded

The compromise began when attackers obtained an access key associated with Fireworks AI, which Cursor relies on for AI model training and inference operations. Using this key, the threat actors accessed a read-only storage bucket that Cursor maintained to exchange data with Fireworks.

The exposed bucket held historical AI coding-agent traces, predominantly from 2025, collected from users who had not activated Privacy Mode and whose data was eligible for model training purposes. Critically, some of these traces contained unredacted API credentials, granting attackers a pathway to compromise additional Cursor systems.

Cursor stated that fewer than 0.05% of its user base had coding-agent traces exposed, while fewer than 0.01% of customers experienced API credential compromise. Leveraging the stolen credentials, the attackers accessed multiple Cursor software-as-a-service platforms and obtained:

  • Archived, non-production Cursor source code repositories containing some business contact information
  • Partial customer support records and email exchanges
  • Product usage data and customer contact information
  • Summaries of certain AI coding-agent conversations
  • Credentials that customers had previously provided to Cursor's support personnel

Cursor's monitoring infrastructure identified the unauthorized access, prompting the company to revoke the Fireworks access key, refresh exposed credentials, and conduct a broader investigation. The company temporarily disabled certain internal and product services as a precautionary measure.

By that time, however, the attackers had already secured a credential permitting them to publish extensions to Cursor's marketplace. On October 29, they leveraged this stolen credential to upload malicious variants of the Claude Code extension.

On October 29, using a stolen credential from the exposed data from our internal system, the attacker placed malicious, unauthorized copies of Anthropic's Claude Code extension directly on the cache used to serve the Cursor VS Code extension marketplace

Cursor

The attack did not compromise Anthropic's legitimate extension or Open VSX, the extension repository that Cursor uses. The malicious builds, identified as versions 2.1.312 and 2.1.313, remained available for approximately 11 hours before Cursor removed them.

Malware Capabilities and Impact

The attackers modified the legitimate Claude Code extension to include malicious JavaScript code designed to deploy persistent SSH backdoors on developer machines. The payload gathered system information, installed an attacker-controlled SSH key, attempted to activate SSH access, and established persistence mechanisms to retain remote access.

Security researchers identified implementation flaws in the malware that likely prevented the backdoor from functioning correctly on many Windows and macOS systems where SSH was not pre-enabled. Cursor estimates that as many as 5,000 users who downloaded affected extension versions during the exposure window may have received the malicious builds.

Cursor stated that its production infrastructure, official application releases, customer code repositories, and code and prompts belonging to users with Privacy Mode enabled were not compromised. The company has since migrated all inference traffic away from Fireworks, revoked and refreshed exposed credentials, rebuilt its extension marketplace, and redeployed its production infrastructure as a safeguard.

Black October Campaign

The Cursor incident is part of a broader series of attacks targeting AI developer platforms that cybersecurity firm Wiz has labeled "Black October". These campaigns have focused on development infrastructure to extract authentication credentials, particularly those granting access to AI models.

Threat actors have weaponized stolen credentials for unauthorized AI inference operations, generating substantial financial damage for targeted organizations. Some incidents have also been connected to attempts at AI model distillation. The campaign has affected Railway, Northflank, Lightning AI, OpenCode, and Warp, among other platforms.

At Railway, attackers stole customer authentication secrets and used them to generate approximately $740,000 in fraudulent AI charges. OpenCode similarly reported exposed API keys and unauthorized AI consumption.

Fireworks AI itself disclosed multiple security incidents in October, including one involving exposed API keys and tokens and another in which attackers gained access to development machines through a compromised third-party cloud provider. The Fireworks breaches also affected other customers, including Notion, which recently disclosed that AI chat transcripts from a limited number of workspaces were exposed.