Junious Media NetworkMonday, November 16, 2026
Juni/Tech
Advertisement
Security

Cursor Confirms Breach Led to Malicious Claude Code Extension

Attackers stole Fireworks AI credentials and used them to publish corrupted Claude Code extensions, exposing up to 5,000 developers to malware.

Key points

  1. Attackers used a stolen Fireworks AI access key to reach Cursor's internal systems and publish two corrupted Claude Code extension builds on October 29.
  2. The malicious extensions, builds 2.1.312 and 2.1.313, were available for about 11 hours and may have reached up to 5,000 developers.
  3. The malware was designed to install persistent SSH backdoors, though coding errors likely prevented it from working on many Windows and macOS machines.
  4. Cursor moved all AI processing traffic away from Fireworks, rebuilt its extension marketplace, and redeployed its production environment after the breach.

Cursor confirmed Friday that a security breach let hackers push corrupted software through its extension store. The company's incident report traces the intrusion to a stolen access key at Fireworks AI, a supplier it relied on for processing AI requests. Using that key as a starting point, the attackers worked their way into Cursor's internal systems and eventually published fake versions of Anthropic's Claude Code extension.

The corrupted builds were live for roughly 11 hours before Cursor took them down, according to BleepingComputer's coverage of the incident. Up to 5,000 developers who updated the extension during that window may have received the bad files. Cursor is a widely used code editor that helps software developers write and fix code with AI assistance. It runs on the same open-source base as Microsoft's Visual Studio Code. An extension marketplace is a library of add-on tools users can install to expand what their editor does, making it a high-value target once an attacker gains publishing access.

How the Cursor and Claude Code Breach Unfolded

The attack started at Fireworks AI. Hackers obtained an access key that Cursor had stored with that provider. They used it to copy files from a read-only storage bucket Cursor shared with Fireworks for model training work. The bucket held records from past AI coding sessions, most of them from 2025. Only users who had not switched on Privacy Mode had session data stored there.

Some of those session records still contained API credentials in plain text. The attackers pulled those credentials and used them to reach deeper parts of Cursor's infrastructure. That path eventually produced a key with permission to publish software directly to the Cursor extension store.

Advertisement

On October 29, the hackers submitted two altered versions of the Claude Code extension, tagged as builds 2.1.312 and 2.1.313, as BleepingComputer reported. Those files sat in the store for about 11 hours before they were taken down.

Advertisement

What the Malicious Files Were Built to Do

The altered extensions carried JavaScript designed to open a persistent backdoor using SSH. SSH, short for Secure Shell, is a standard protocol for encrypted remote connections. A backdoor built on it allows an outside party to log in to a machine repeatedly without needing new credentials each time.

The malware also gathered basic facts about the host machine, planted an attacker-controlled SSH key, tried to switch on SSH access where it was not already running, and set up persistence so the entry point would survive a reboot.

Security researchers examined the code and found programming mistakes that likely stopped the backdoor from activating on many Windows and macOS machines where SSH was not already running. Those errors reduced the risk for some users, though not for all. Cursor confirmed that Anthropic's real Claude Code extension was not modified, and the Open VSX registry that Cursor normally draws from was also untouched.

Who Was Affected and What Data Was Reached

Cursor reported that fewer than 0.05% of its users had session records reached through the Fireworks storage bucket. Fewer than 0.01% of customers had API credentials taken from those records. The percentages are small, but a widely used tool can still translate small fractions into a meaningful number of people.

Advertisement

Attackers also reached other internal material at Cursor. That included archived non-production source code, limited customer support messages and email exchanges, product usage statistics, and credentials some customers had previously sent to Cursor's support team. The company confirmed that production systems, official app releases, user code, and the data of anyone with Privacy Mode switched on were not affected.

How Cursor Responded to the Attack

Cursor's monitoring tools detected the unauthorized access. Engineers immediately canceled the Fireworks key, started rotating all exposed credentials, and briefly shut down some internal and product services while they assessed the full scope of the intrusion. The company also launched a broader investigation to map how far the attackers had moved.

After containing the immediate problem, Cursor shifted all AI processing traffic away from Fireworks entirely, BleepingComputer reported. The company rebuilt its extension marketplace from scratch and redeployed its production environment. Those steps were meant to ensure nothing from the affected infrastructure remained in active use.

Advertisement

Cursor advises any developer who installed or updated Claude Code around October 29 to follow the company's published remediation steps. Rotating API keys, tokens, and any other active credentials from that period is the most urgent action.

Part of a Broader Campaign Against AI Developer Tools

The Cursor incident fits into a wider attack campaign. Security firm Wiz labeled the stretch "Black October," a series of credential-theft operations targeting companies that build tools and infrastructure for AI developers. Stolen credentials were used for unauthorized AI inference, running up large fraudulent charges on victims' accounts. Some incidents were also tied to AI model distillation, a technique for copying a model's behavior into a smaller system.

Advertisement

Other platforms hit in the same period include Railway, Northflank, Lightning AI, OpenCode, and Warp. At Railway, stolen secrets were later used to generate roughly $740,000 in fraudulent AI charges. OpenCode reported stolen API keys and unauthorized model use around the same time.

Fireworks AI also disclosed several separate incidents in October. One involved exposed API keys and tokens. A second allowed attackers to reach development machines through a compromised outside cloud provider. Those events had downstream effects on other Fireworks customers: Notion recently disclosed that AI chat logs from a small number of its workspaces were exposed as a result of the Fireworks incidents.

What Developers Should Do Now

For anyone who may have been affected, Cursor's official remediation guide is the clearest starting point. Developers should also confirm that any currently installed version of the extension was obtained after Cursor rebuilt its marketplace. Any credential that was active in the affected environment during the exposure window should be rotated without delay.

This incident points to a structural risk in AI development tooling. When multiple products share infrastructure from a common provider, a single stolen key can ripple across the whole ecosystem. Reviewing what data is shared with third-party AI providers, keeping credentials scoped as narrowly as possible, and watching for unusual outbound network activity are practical steps that reduce exposure if a shared supplier is hit again.

Advertisement

Sources