Due to a security incident at AI infrastructure provider Fireworks AI, attackers gained access to Cursor’s internal systems. Using stolen login credentials, they then distributed malicious versions of Anthropic’s “Claude Code” extension.
According to a security update from Cursor, the attackers obtained an access key stored at Fireworks AI. Using this key, they copied historical conversations with AI coding assistants from a storage location. These conversations were primarily from 2025 and were used for training models.
Some of these conversations contained login credentials that had not been deleted, allowing the attackers to access various SaaS tools used by Cursor. As a result, archived source code repositories, customer support data, email correspondence, and additional login credentials were exposed.
Cursor estimates that conversations with AI coding assistants involving less than 0.05 percent of users were exposed. Additionally, API keys belonging to less than 0.01 percent of customers were compromised.
5,000 potential downloads
On October 29, the attackers used other stolen credentials to cache modified versions of Anthropic’s “Claude Code” extension. This cache is used for Cursor’s VS Code extension marketplace. The compromised versions, 2.1.312 and 2.1.313, remained available for approximately 11 hours. Up to 5,000 users may have downloaded these versions.
As previously reported by BleepingComputer, the malicious extension contained JavaScript code designed to collect system information. It was also intended to install an attacker-controlled SSH key and gain persistent access to the system. However, due to programming errors, the backdoor likely did not work on many Windows and macOS systems. In particular, it did not work on systems where SSH was not already enabled.
Cursor emphasizes that the legitimate extension from Anthropic and the Open VSX marketplace have not been compromised. The company further states that its production infrastructure, official application versions, and customer code repositories remained unaffected.
A broader wave of attacks
Cursor has since removed the malicious extensions and rebuilt the marketplace. In addition, it replaced exposed login credentials and rerouted all traffic for AI model inference away from Fireworks AI. According to the company, the attackers’ last observed activity was blocked on November 1.
This incident is part of a broader wave of attacks on AI developer platforms. Security firm Wiz dubbed this wave “Black October.” Railway, OpenCode, and Warp were also affected by security incidents. The attackers targeted login credentials that grant access to AI services. At Railway, stolen login credentials reportedly led to about $740,000 in fraudulent AI usage charges.