Back to Flash News

Cursor investigation confirms Claude Code plugin poisoning stemmed from Fireworks key leak

On 2026-11-13, Cursor announced the results of its investigation into the earlier poisoning incident involving the Claude Code plugin. After a two-week investigation, the company said the source of the attack was its AI service provider Fireworks.

Cursor said the attacker obtained the access key that the company had granted to Fireworks, and that the key was leaked by Fireworks. The attacker used the key to copy historical Coding Agent run records from a shared read-only storage bucket. These records mainly came from user sessions in 2025 with privacy mode turned off and were originally used for model training. Some records contained API keys that had not been redacted, which enabled the attacker to access some SaaS tools used by Cursor and obtain archived non-production code, customer service emails and conversations, product usage data, some AI conversation summaries, and credentials submitted by users to customer service.

On October 29, the attacker used credentials stolen from internal systems to place a malicious, unauthorized version of the Claude Code extension directly into the cache of the Cursor VS Code extension marketplace. The malicious versions were 2.1.312 and 2.1.313, and they existed for about 11 hours. Up to 5,000 users requesting the specific cached version may have received it. The upstream Open VSX was not tampered with, Anthropic did not release this copy, and its product also had no related vulnerability.

Cursor disclosed that the Coding Agent run records involved fewer than 0.05% of users, and fewer than 0.01% of users had API credentials exposed. Sessions with privacy mode enabled, customer code repositories, and production systems were not affected. The company said the last time it detected attacker activity was on November 1, when it was successfully intercepted. Affected users have now been notified, Fireworks has been fully stopped from handling inference requests, its data access privileges have been revoked, internal credentials have been rotated, and production infrastructure has been redeployed.

Original source:

  1. Title: An update on our recent security incident
    Publisher: Cursor
    Publication date: 2026-11-13
    URL: cursor.com/blog/security-incident-update
    Source type: official company blog announcement
    Verification status: the full official text was actually read for this report, verification date 2026-11-13
    Supporting content: the attack began with the Cursor-issued key leaked by Fireworks, copying of historical Coding Agent records, unredacted API credentials, access to internal SaaS tools, the October 29 malicious extension deployment, up to 5,000 users, an 11-hour window, the impact scope, and the company’s response measures

ABAB AI Insight

Cursor’s official update traces the incident chain back from the malicious extension discovered on October 29 to the access key leaked by Fireworks. Using that key, the attacker copied historical Coding Agent records used for model training from a shared read-only storage location, mainly from 2025 sessions without privacy mode enabled, then extracted unredacted credentials from them and further accessed some of Cursor’s internal SaaS tools, obtaining archived non-production code, customer service data, and some conversation summaries.

On October 29, the attacker used stolen credentials to directly tamper with the extension marketplace cache and deploy malicious versions of Claude Code 2.1.312 and 2.1.313 that had not been released by Anthropic. The window was about 11 hours, and the maximum number of potentially affected users was 5,000. Cursor made clear that Open VSX and Anthropic’s official channels were not affected, and that production systems, privacy-mode sessions, and customer code repositories were not impacted.

The company has now moved all inference traffic away from Fireworks, revoked the relevant access, rotated credentials, and redeployed infrastructure. The last intercepted attacker activity occurred on November 1. This incident shows that leaked access keys from a third-party AI service provider can, through a training-data storage path, propagate to internal tools and extension distribution caches.

ABAB News · Laws of Cognition

  1. An access key leaked by an AI service provider can be used to copy historical training data in shared storage.

  2. Unredacted API credentials can propagate from historical session records to access to internal SaaS tools.

  3. An extension marketplace cache can be used to deploy malicious versions not released upstream, with an impact window measured in hours.