Leaked Fireworks Key Led to Cursor Backdoor
Cursor said on Nov. 13 that the tampered Claude Code extension in its marketplace on Oct. 29 traces to an access key exposed at Fireworks AI. The attacker copied coding-agent traces, used unmasked credentials in them to reach Cursor's SaaS tools, then planted the builds. Up to 5,000 users may have received them.
Cursor said on Friday, Nov. 13, that the backdoored Claude Code extension that reached its marketplace on Oct. 29 traces back to an access key exposed at Fireworks AI, the training and inference provider it used. The attacker used that key to copy old coding-agent traces, mined them for credentials, and climbed from there into Cursor's own systems.

Key takeaways
- The breach began with a key Cursor had issued to Fireworks, which was exposed on the Fireworks side and used to copy data from a shared read-only storage bucket.
- Traces in that bucket held unmasked credentials; fewer than 0.01% of customers had API credentials exposed, and fewer than 0.05% of non-Privacy Mode users had traces exposed.
- A stolen credential let the attacker plant tampered Claude Code builds on the marketplace cache. Cursor has moved all inference off Fireworks and says the incident is contained.
How the Cursor breach started at Fireworks
In its incident update, the SpaceX-owned code editor company said the attacker obtained an access key Cursor had issued to Fireworks. The key was exposed on Fireworks' side, and the attacker used it to copy data from a read-only storage bucket the two companies used to share data.
The bucket held a limited set of coding-agent traces from sessions that were not in Privacy Mode. Cursor said most of the data dates from 2025 and was licensed for model training.
Some of those traces contained credentials that had not been masked. The attacker used them to reach several software-as-a-service tools Cursor uses.
What data the attacker reached
Through those tools, Cursor said, the attacker accessed archived, non-production source code repositories, some with business contact details. Also reached were limited customer support data and email conversations, product usage metrics and contact information, some coding-agent conversation summaries, and some credentials users had sent to Cursor's support team.
Cursor said it has identified and individually notified every affected user. Users who did not receive a notice were not in the exposed traces, according to the company.
From stolen credential to backdoored extension
On Oct. 29, the attacker used a credential taken from the exposed data to place tampered copies of Anthropic's Claude Code extension directly on the cache that serves Cursor's VS Code extension marketplace. Cursor said the copies were not published by Anthropic and did not stem from a flaw in an Anthropic product. Open VSX, the marketplace's upstream source, never carried the malware.
Up to 5,000 users who requested specific cached versions, 2.1.312 and 2.1.313, may have received the malicious builds during an exposure of about 11 hours. BleepingComputer reported that the code tried to plant persistent SSH backdoors, though researchers found coding errors likely kept it from working on many machines without SSH already enabled.
Cursor warned that simply removing the extension may not be enough and pointed affected users to its original security notice, which lists indicators of compromise and a check script.
What Cursor says was not affected
The company said prompts and code from Privacy Mode users, live traffic, production systems and customer code repositories were untouched. The Cursor app and its official updates ship from a separate code hosting system the attacker never reached, and release signing keys were not exposed.
Cursor also said it verified every other extension in the marketplace as authentic. Other products, including its cloud agents and the Grok bot, were not affected.
Context: a month of attacks on AI platforms
The disclosure links two threads of what cloud security firm Wiz calls "Black October." Wiz counted at least nine AI developer platforms disclosing incidents in October, with Fireworks first, and has kept adding cases into November. Wiz says the attacks share a hunt for stored credentials that grant AI access.
Cursor is not the only downstream victim. Wiz's running log notes that Notion gave notice of downstream impact from the Fireworks incident, which Fireworks said on Nov. 9 reached billing records, some datasets and incomplete model weights.
What's next
Cursor said the last observed attacker activity, on Nov. 1, was blocked. It has disabled the Fireworks key, cut access to shared datasets, rotated credentials across internal systems, added secret scanning to all code repositories, tightened network restrictions and redeployed production infrastructure. External forensic experts are assisting, and Cursor said it will update the post if anything material changes.
FAQ
Was Anthropic's Claude Code compromised?
No. Cursor said the tampered copies were placed on its own marketplace cache and did not come from a vulnerability in an Anthropic product.
How do I know if my coding-agent traces were exposed?
Cursor said it individually notified every affected user. If you did not get a notice, the company says your data was not in the exposed traces.
I installed version 2.1.312 or 2.1.313 in Cursor. What now?
Follow the steps in Cursor's Oct. 30 security notice. Uninstalling alone may leave persistence behind.
References
- An update on our recent security incident — Cursor (Nov. 13, 2026)
- Security notice: malicious version of Anthropic.claude-code — Cursor (Oct. 30, 2026)
- Cursor confirms security breach led to malicious Claude Code extension — BleepingComputer (Nov. 14, 2026)
- "Black October" for AI Developer Platforms — Wiz (Oct. 30, 2026, updated)
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.