Cursor Breach: Backdoored Claude Code Hit 5,000 Devs Cursor confirmed that a compromised Fireworks AI inference vendor key let attackers serve backdoored builds of Anthropic's official Claude Code extension — versions 2.1.312 and 2.1.313 of anthropic.claude-code — to up to 5,000 developers for roughly eleven hours on October 29, according to Cursor. The malicious builds added a 66-line JavaScript loader that attempted to plant a persistent SSH backdoor via a key labeled svc-telemetry, a reverse tunnel, and a macOS LaunchAgent or Windows scheduled task retrying every 60 seconds; Cursor said the attack largely failed because enabling SSH requires administrator privileges and the code contained errors. Cursor warned that dropped artifacts live outside the extension folder, so exposed developers must run its cleanup scripts rather than simply uninstalling or updating, and research cited in the report found 31,088 distinct secrets in .claude/ files. Cursor confirmed last week that a compromised inference vendor key let attackers serve a backdoored version of Anthropic’s official Claude Code extension to up to 5,000 developers for eleven hours. The malicious builds — versions 2.1.312 and 2.1.313 of anthropic.claude-code — sat in Cursor’s VS Code marketplace cache on October 29, containing a 66-line JavaScript loader that attempted to plant a persistent SSH backdoor on developer machines. This was not a compromised build pipeline. The attack started with something most security teams haven’t modeled yet: an AI training vendor that exposed a client key, which led directly to developer credentials embedded in historical AI session logs. How a Vendor Key Became a Marketplace Backdoor The chain is worth understanding step by step, because each link represents a failure mode that will appear again. Cursor had issued an access key to Fireworks AI https://fireworks.ai/security/bulletins/october-2026-security-incident , one of its inference and model training vendors. Fireworks experienced a breach in early October — a multi-stage incident involving stolen internal credentials — and in the process exposed the key Cursor had given it. That key provided read access to a storage bucket holding historical coding-agent traces: session logs from Cursor users who had not enabled Privacy Mode, shared with Fireworks for model training purposes. Some of those traces contained unredacted API keys and tokens that developers had pasted directly into AI chat. The attacker used those credentials to pivot into Cursor’s internal SaaS tools — accessing archived source code, customer support data, product metrics, and credentials submitted to Cursor’s support team. On October 29, using a credential obtained through that internal access, the attacker wrote unauthorized builds directly to the cache serving Cursor’s extension marketplace. Versions 2.1.312 and 2.1.313 of the official Claude Code extension were live for roughly eleven hours before Cursor detected and pulled them. What the Payload Did and Why Most Machines Were Spared The malicious extension inserted 66 lines of JavaScript before the legitimate extension code, running automatically at module load with no user interaction needed beyond installation or auto-update. The loader collected host and user information, attempted to add an SSH key labeled svc-telemetry to ~/.ssh/authorized keys , enable the local SSH service, and open a reverse tunnel to an attacker-controlled server. For persistence, it dropped a LaunchAgent on macOS or a scheduled task on Windows, with retry scripts that attempted to re-establish the tunnel every 60 seconds. In practice, the attack largely failed. Enabling the SSH service requires administrator privileges, and most macOS and Windows machines don’t run Cursor with admin rights. Cursor confirmed coding errors in the malicious code further limited its success. But there is a critical catch: the artifacts the malicious extension drops — the SSH key, the persistence scripts, the LaunchAgent or scheduled task — live outside the extension folder. Uninstalling the extension does not remove them. Developers who may have been exposed must run Cursor’s cleanup scripts explicitly, not just update or reinstall. The Attack Surface Security Teams Missed: Session Traces as Credential Vaults The more important story here is not the SSH backdoor. It’s the intermediate step: unredacted credentials in AI session logs. Developers routinely paste API keys, database connection strings, and tokens directly into AI coding assistants. Those sessions, when shared with inference or training vendors without Privacy Mode enabled, create a credential cache that most security models don’t account for. Research from BleepingComputer’s coverage https://www.bleepingcomputer.com/news/security/cursor-confirms-security-breach-led-to-malicious-claude-code-extension/ and GitGuardian found 31,088 distinct secrets in .claude/settings.local.json files across 23,276 public GitHub repositories — 5,343 of them still valid. Boris Cherny, who leads Claude Code at Anthropic, acknowledged the tool does not yet do secrets scrubbing, with built-in scanning under exploration. This is not a Cursor-specific problem. Any AI coding tool that shares session data with training or inference vendors creates this surface. The Cursor incident demonstrates what happens when a vendor key to that data is compromised: the attacker gets not just LLM usage logs, but potentially a credential dump of everything developers have pasted into AI chat over months or years. What Affected Developers Should Do Now If you were using Cursor on October 29 and may have installed or auto-updated Claude Code that day, take these steps: - Update the Claude Code extension to version 2.1.314 or later - Run Cursor’s official cleanup script https://cursor.com/blog/security-incident-update — uninstalling alone is not sufficient - Check ~/.ssh/authorized keys for any entry labeled svc-telemetry and remove it - Check ~/Library/LaunchAgents/ on macOS or Windows Task Scheduler for suspicious persistent entries - If you found evidence of an active SSH tunnel, treat the machine as compromised and rebuild Marketplace Trust Is Not Build Provenance Going forward, enable Privacy Mode in Cursor and equivalent settings in other AI coding tools. Treat AI chat sessions as insecure channels — do not paste real credentials into them, and rotate any credentials you may have shared with AI tools in the past year. Apply the same scrutiny to AI tool marketplace extensions that you would to any third-party dependency. The attacker in this incident bypassed Open VSX entirely by writing directly to Cursor’s marketplace cache, leaving the official publisher identity unchanged. “This came from the official marketplace” is not equivalent to “this build was verified against the upstream source.” As AI coding tools become deeper infrastructure for development workflows, the security assumptions applied to them need to match the access they are being granted https://www.securityweek.com/malicious-claude-code-extension-served-in-cursor-supply-chain-attack/ .