Cursor confirms breach: stolen Fireworks key, customer data exposed
Cursor's post-mortem ties the malicious Claude Code extension to an access key stolen from Fireworks AI, plus exposure of archived source, support data and some customer credentials.
AI code editor Cursor has published a post-mortem confirming that the malicious Claude Code extension served from its marketplace on 29 October was one symptom of a broader intrusion that also exposed customer data. Per Cursor's incident update, the root cause was a stolen access key that Cursor had issued to Fireworks AI, a third-party model provider. If you used Cursor without Privacy Mode in 2025, or sent credentials to Cursor support, check for a notification and rotate anything that was in scope.
Root cause
Cursor states plainly: "An attacker obtained an access key we had issued to Fireworks." The key was exposed on Fireworks' side; the attacker used it to copy data from a read-only storage bucket Cursor used to share data with Fireworks. Some of that data held unredacted credentials, which the attacker then used to reach a handful of Cursor's SaaS tools and move laterally. That lateral access — not a flaw in the extension marketplace itself — is what let the attacker publish the tampered extension builds on 29 October.
What was exposed
By Cursor's account, the attacker accessed:
- Coding-agent traces, mostly from 2025, from sessions not in Privacy Mode.
- Archived non-production source-code repositories, some containing business contact information.
- Limited customer support data and email conversations.
- Product usage metrics and contact details, plus summaries of some coding-agent conversations.
- Credentials users had sent to Cursor's support team.
Cursor puts the blast radius in percentages: fewer than 0.01% of customers had API credentials exposed, and for non–Privacy Mode users, traces were exposed for fewer than 0.05%. Production infrastructure, official application releases, customer code repositories, and the code and prompts of Privacy Mode users were not affected, the company says.
The extension, again
The marketplace half is unchanged from the earlier notice: malicious copies of anthropic.claude-code versions 2.1.312 and 2.1.313 sat on the cache serving Cursor's VS Code marketplace for about 11 hours on 29 October. At most 5,000 users who requested those cached versions that day could have received the tampered build, which attempted to install a persistent SSH backdoor. Anthropic did not publish the builds, and the Open VSX registry was never affected. Cursor says the last observed attacker activity was on 1 November, and it was blocked.
Action checklist
- Check for a Cursor notification. Cursor says it contacted affected customers directly and that team admins received lists of specific keys to rotate. If you were not notified, Cursor says your data was not in the affected traces.
- Rotate credentials you ever sent to Cursor support. Any API key, token, or secret pasted into a support ticket should be treated as exposed and rotated.
- If you installed 2.1.312 or 2.1.313 from the Cursor marketplace during the window, follow the steps in Cursor's security notice. The company warns: "Uninstalling the extension alone may not be enough" — the backdoor's SSH key and persistence survive a plain uninstall.
- Rotate anything that lived in a 2025 non–Privacy Mode session or an archived repo. Hard-coded secrets in old agent traces or non-production repos are the credentials most likely to have ridden along.
- Audit for the attacker SSH key on any developer machine that ran the malicious extension, per the IOCs in Cursor's 30 October notice.
Context
The pivot here is the same one that is defining this autumn's run of incidents: a key issued to a third-party AI provider becomes the entry point, and a read-only bucket full of unredacted secrets turns a single exposed credential into lateral movement. Security firm Wiz groups the surrounding activity under a "Black October" wave against AI developer platforms; reporting at BleepingComputer and BankInfoSecurity ties the same Fireworks AI exposure to Notion and names Railway, Northflank, Lightning AI, OpenCode and Warp among others caught in it (eSecurityPlanet). The lesson is the one Hacker Posts flagged after the Warp agent-log breach: the keys you hand a vendor are keys into you, and a storage bucket that stores secrets in the clear is a breach waiting for one leaked token.