Cursor confirmed on November 14, 2026, that an access key it had issued to Fireworks AI was stolen, causing malicious Claude Code extension versions 2.1.312 and 2.1.313 to be uploaded to the marketplace on October 29. About 5,000 developers may have installed versions containing a persistent SSH backdoor.
Reconstruction of Events
According to an investigation report published by Cursor, the attack began with the leak of an access key stored by Fireworks AI. The attacker used the key to copy historical AI coding-agent traces from a read-only storage bucket. These traces mainly came from user data from 2025 when Privacy Mode was not enabled, and some contained unredacted API credentials. The credentials were then used to access Cursor's internal systems.
The report shows that fewer than 0.05% of user traces were exposed and fewer than 0.01% of customer API credentials were affected. The attacker used the credentials to steal non-production code repositories, limited customer support information, product usage metrics, and some AI conversation summaries. Cursor's monitoring system detected anomalies, revoked the key, and launched an investigation, but the attacker had already obtained extension marketplace distribution credentials.
On October 29, the malicious extension was listed on Cursor's VS Code extension marketplace cache and was removed after about 11 hours. The extension modified Anthropic's legitimate Claude Code code and injected JavaScript that attempted to install an SSH backdoor on the host, including collecting system information, adding attacker SSH keys, and configuring persistence. Cursor said Anthropic's official extension and the Open VSX marketplace were not affected.
Mechanism Breakdown
The incident formed a three-stage intrusion chain: first, read-only access was obtained through the storage bucket of third-party vendor Fireworks AI; then plaintext credentials were extracted to enter Cursor's internal systems; finally, distribution credentials were used to compromise the extension marketplace. The malicious extension responded normally to developer requests while installing the backdoor, attempting to enable SSH on Windows and macOS systems.
Security researchers found flaws in the code that could cause the backdoor to fail on systems where SSH was not previously enabled. Cursor has moved all inference traffic away from Fireworks, rebuilt the extension marketplace, and redeployed production infrastructure.
Industry Impact
This incident, together with multiple attacks on AI development platforms in October, has been described as part of "Black October." Similar attacks targeted companies such as Railway and Northflank, all aimed at stealing AI model access credentials. Affected Cursor users need to clean their systems according to company guidance.
The incident highlights AI coding tools as high-value attack surfaces: credential leaks can directly compromise developers' local environments without breaching the model. Fireworks customers such as Notion also reported related exposure.
Strategic Assessment
[Analysis] Based on the available facts, the AI toolchain's dependence on third-party vendors amplifies the cascading effects of a single key leak, with developer endpoints becoming the ultimate target. Enterprises may need to reassess the trust model of extension marketplaces and credential storage methods to reduce similar supply chain contamination risks.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接