Cursor 資料外洩事件:Claude Code 擴充功能在 Privacy Mode 開啟與關閉時的差異
Cursor Data Breach Claude Code Extension: Privacy Mode vs Off
Cursor 表示,惡意 Claude Code 擴充功能副本曾經其市集快取分發;Privacy Mode 內的提示詞和程式碼未受影響,外洩的編碼 Agent 軌跡來自未啟用該模式的工作階段。
In the Cursor data breach, a leaked Fireworks access key led to a malicious copy of the Claude Code extension being served from Cursor's extension marketplace cache. Cursor's update of 13 November 2026 says at most 5,000 users who requested specific cached versions on 29 October may have received it, and that it was available for about 11 hours. The verdict is clear: Privacy Mode on meant prompts and code stayed unaffected; the exposed data came from sessions without it. Cursor believes the incident is contained, so the live task is cleanup for anyone who installed Anthropic.claude-code 2.1.312 or 2.1.313 on 29 October.
TL;DR
- The attack began at Fireworks, a vendor, not in Anthropic's product. Anthropic did not publish the malicious copy, according to Cursor.
- Tampered versions were 2.1.312 and 2.1.313 for Windows and macOS, served between about 2:27am and 1:35pm PT on 29 October.
- Fewer than 0.01% of customers had API credentials exposed, and fewer than 0.05% of non-Privacy-Mode users had coding agent traces exposed, per Cursor.
- Privacy Mode prompts and code, production systems, customer repositories and the Cursor app itself were not affected.
- If you installed an affected version, check for an SSH key labelled svc-telemetry; uninstalling alone may not be enough.
How did the Cursor data breach reach a Claude Code extension?
The chain started with a key: an attacker obtained an access key Cursor had issued to Fireworks, an AI inference and training vendor, after Fireworks exposed it. Fireworks' own security bulletin describes a separate incident at a third-party cloud provider hosting its research machines on 23 October, followed by unauthorised access to development machines on 24 October.
With the key, the attacker copied data from a read-only storage bucket that Cursor shared with Fireworks. It held limited coding agent traces, mostly from 2025, from sessions outside Privacy Mode that were permitted for model training. A small number of those traces contained credentials that had not been redacted.
From those credentials the attacker reached select SaaS tools used by Cursor. Cursor reports access to archived non-production source repositories, limited support data, usage metrics, some agent-conversation summaries and credentials users had sent to support. On 29 October, a stolen credential let the attacker place malicious copies of the Claude Code extension on the cache that serves Cursor's marketplace. Cursor says the upstream Open VSX marketplace never held the malware. BleepingComputer first reported the tampered extension on 29 October; its coverage of the full report confirms the Fireworks-key origin.
What is the timeline of the incident?
| Date (2026) | Event | Source |
|---|---|---|
| 23 Oct | Incident at a third-party cloud provider hosting Fireworks R&D machines | Fireworks |
| 24 Oct | Unauthorised party reaches Fireworks development machines | Fireworks |
| 25 Oct | Fireworks reports no further successful access to its customer platform | Fireworks |
| 29 Oct, 2:27am to 1:35pm PT | Malicious extension copies served from Cursor's cache | Cursor notice |
| 29 Oct | Takedown and public security notice | Cursor notice |
| 30 Oct | Notice updated with indicators of compromise | Cursor notice |
| 1 Nov | Last attacker activity observed, and blocked | Cursor update |
| 9 Nov | Fireworks bulletin last updated | Fireworks |
| 13 Nov | Cursor publishes its full investigation update | Cursor update |
Who was affected, and how many people is that?
Cursor's report splits the exposure into three separate populations.
- Extension downloads: at most 5,000 users who requested the specific cached versions on 29 October.
- API credentials: fewer than 0.01% of customers had credentials exposed through the traces.
- Coding agent traces: fewer than 0.05% of non-Privacy-Mode users had traces exposed, and Cursor notified each of them directly.
The 5,000 figure is a ceiling, not an infection count; the malicious code often failed to obtain access, as the next section explains.
What did the malicious extension actually do?
According to Cursor's 29 October notice, the payload collected host details: hostname, username, OS, architecture, home directory, environment variable names, Git email and public IP. It then tried to add an attacker-controlled SSH key to authorized_keys, enable the local SSH service and open a reverse SSH tunnel. Persistence used a LaunchAgent on macOS and a scheduled task on Windows, with background scripts retrying the tunnel about every 60 seconds - a chain independent write-ups traced the same way (Cybersecurity News).
The backdoor frequently failed end to end: enabling the local SSH service needs administrator rights, which the extension lacks on most macOS and Windows hosts, so the attacker usually could not log in through the tunnel. The key, persistence scripts and host report were still delivered, which is why cleanup matters even where no login occurred.
Why did this chain work, and what does it show?
One exposed vendor key unlocked a data set that contained live credentials: traces meant for model training held secrets nobody had stripped, turning a vendor leak into an internal breach. The same pattern appears in the Chainalysis breach of password hashes and tokens.
Two controls stand out: Privacy Mode kept sensitive prompts and code out of the shared bucket, and credential redaction would have limited the damage for everyone else. Neither company has named the attacker. Fireworks links its incident to a financially motivated group engaged in what it calls LLMJacking, but Cursor does not attribute its own incident, and this article does not either.
The blame sits with a vendor supply chain, not with Anthropic's product; anyone running agents with broad access should store Claude Code secrets in a vault, since secrets that never appear in a session cannot leak into a trace.
What should defenders do now?
Use only the vendor-published steps: Cursor's notice includes indicators of compromise, a validation script and cleanup guidance.
| Step | Who runs it | Where published |
|---|---|---|
| Search authorized_keys for the key label svc-telemetry (grep on macOS, PowerShell on Windows) | Anyone who had 2.1.312 or 2.1.313 on 29 October | Cursor notice |
| Run the vendor validation script and remove the SSH key, persistence mechanisms and tunnel processes | Security or IT teams, or the individual user | Cursor notice |
| Isolate affected hosts, consider a rebuild, and rotate credentials and sessions from a trusted system | Security teams | Cursor notice |
Cursor's own response included moving all inference traffic off Fireworks, rotating every exposed and internal credential, rebuilding the marketplace and engaging external forensics experts (Unite.AI recaps the full list). Fireworks has notified law enforcement and engaged Palo Alto Networks Unit 42.
Related reading: a similar race between patch and exploit in OpenSSH, Cursor's hosted cloud agents and Claude Code 2.1.323 tool calls.
FAQ
Q: Was the Claude Code extension itself vulnerable?
A: No. Cursor states this was not an Anthropic product vulnerability and Anthropic did not publish the copy; the tampered files were placed on Cursor's marketplace cache.
Q: Which versions were malicious?
A: Anthropic.claude-code 2.1.312 and 2.1.313 for Windows and macOS, available for about 11 hours on 29 October 2026.
Q: Does Privacy Mode protect me in this incident?
A: Cursor says Privacy Mode prompts and code were not affected; the exposed traces came from sessions outside Privacy Mode, permitted for model training.
Q: Is uninstalling the extension enough?
A: Cursor warns it may not be. The payload drops standalone persistence scripts, so check for the svc-telemetry SSH key, remove persistence and tunnels, and rotate credentials.
Q: Is the incident over?
A: Cursor reports the last attacker activity on 1 November was blocked and believes the incident is contained. Neither company has named the attacker.
來源:DEV Community · dev.to