Cursor says Fireworks key leak led to fake Claude Code extension
Cursor's incident report says a key exposed at Fireworks let an attacker steal data and credentials, then push a backdoored Claude Code extension to up to 5,000 users.

- What happened
- Cursor published a detailed account of a breach that began with an access key exposed by its AI vendor Fireworks.
- What’s new
- Cursor now explains how the attacker moved from stolen training traces to its internal tools and then to its extension marketplace cache.
- Why it matters
- Developers who installed the tampered extension versions may still have a backdoor, and Cursor warns that uninstalling the extension may not remove it.
Cursor has published a full account of the security incident behind the malicious copy of Anthropic's Claude Code extension that appeared in its VS Code extension marketplace on 29 October. In the report, dated 13 November, Cursor says the breach began at Fireworks, a third-party provider it used for model training and inference.
From a vendor key to the marketplace
According to Cursor, an attacker got hold of an access key that Cursor had issued to Fireworks and that Fireworks exposed. The attacker used the key to copy the contents of a read-only bucket where Cursor stored data shared with the vendor. That bucket held coding-agent traces, mostly from 2025, from sessions where Privacy Mode was off. Some traces included unredacted credentials. Cursor says traces were exposed for fewer than 0.05% of users and API credentials for less than 0.01% of customers.
Cursor says the attacker used those credentials to reach several of its SaaS tools. From there the attacker took archived non-production source code, some customer support data and emails, usage metrics, contact details, summaries of some agent conversations, and credentials that users had sent to Cursor support. One stolen credential then let the attacker place tampered copies of the extension, versions 2.1.312 and 2.1.313, directly on the cache that serves Cursor's marketplace. Cursor estimates that at most 5,000 users received the malicious build, which stayed live for about 11 hours. Cursor says Anthropic's own extension and the Open VSX marketplace source were never compromised.
What the malware did and what to do
BleepingComputer reports that the modified extension ran JavaScript that tried to install an attacker-controlled SSH key and set up persistent remote access. The outlet adds that security researchers found coding errors that probably kept the backdoor from functioning on many Windows and macOS machines where SSH was not already enabled.
Cursor asks anyone who installed or updated to either tampered version to follow the steps in its 30 October security notice. Cursor warns that removing the extension may not be enough on its own. The company says production systems, official app releases, customer code repositories and Privacy Mode data were unaffected. It has moved all inference off Fireworks, rotated credentials and rebuilt its marketplace. Cursor says the incident is contained and that the attacker's last activity, on 1 November, was blocked.
Part of a wider wave
BleepingComputer places the incident within a wave of attacks on AI developer platforms that security firm Wiz calls Black October. The outlet says the attackers sought credentials for AI model access. Railway, Northflank, Lightning AI, OpenCode and Warp were also hit. At Railway, stolen secrets led to roughly $740,000 in fraudulent AI charges. According to BleepingComputer, Fireworks disclosed several breaches in October, and Notion, another Fireworks customer, said a handful of workspaces had their AI chat transcripts exposed.