Cursor, the company behind the popular AI-powered code editor, has revealed that it was the target of a supply chain security attack resulting from a breach of the AI inference platform Fireworks AI. This breach led to the leakage of authentication data and access keys belonging to Cursor, enabling attackers to compromise parts of its infrastructure and distribute malicious versions of Anthropic's Claude Code extension on the editor's extension marketplace.
Breach Sequence and Attack Vector
Investigations revealed that the breach began after Fireworks AI suffered two consecutive security incidents during October. In the first incident, internal credentials were used to access stored secrets, account information, and billing data. In the second incident, a third-party cloud service provider was compromised, allowing attackers to access Fireworks AI development machines and steal customer data.
During these incidents, the attackers obtained a Cursor access key that was stored at Fireworks AI. This key was used to extract data from a read-only storage unit, which contained unencrypted credentials that opened the door to accessing the Software-as-a-Service (SaaS) tools that Cursor relies on internally.
Affected Data and Nature of the Leak
The access enabled the attackers to reach a disparate set of digital assets and internal data belonging to Cursor, including:
- Non-production and archived source code repositories, some of which contained business contact information.
- Limited customer support data and email correspondence.
- Product usage metrics and customer contact data.
- Summaries of coding agent conversations, in addition to some credentials that users had sent to the support team.
Planting the Malicious Extension and Scope of Damage
Although Cursor blocked the attackers' access immediately upon discovering the unauthorized activity on October 29, the attackers exploited the leaked secrets to plant modified and malicious versions of the Anthropic Claude Code extension (versions 2.1.312 and 2.1.313) directly into the cache used to serve the Cursor VS Code extension marketplace.
Cursor confirmed that Anthropic was not directly breached and that the malicious software was not present in the original Open VSX marketplace source, but was only served to users who requested the cached versions from the marketplace during a window of approximately 11 hours. It is estimated that the maximum number of users affected by the event does not exceed 5,000.
Why This News Matters
This incident highlights the growing risks of supply chain attacks in AI development environments, where a vulnerability in a supporting platform (such as Fireworks AI) can extend to infect widely used development platforms like Cursor. The incident also demonstrates how attackers can exploit extension marketplace caches to pass malicious software packages without needing to compromise the original developer of the extension (Anthropic).
It is worth noting that other companies such as Notion and Uber have also announced that they suffered data leaks linked to the same Fireworks AI platform breach, while Cursor confirmed that the incident has been fully contained and no attacker activity has been observed in its environment since November 1.