One Exposed Vendor Key, 5,000 Poisoned Downloads: Cursor Traces Its Tampered Claude Code Extension to Fireworks AI

On Friday, November 13, Cursor published the full account of a security incident it first disclosed on October 29 — and the chain it describes is a case study in how AI tooling’s supply chain actually fails. The attack did not start with a vulnerability in Cursor’s product, or in Anthropic’s. It started with a single access key at a third-party vendor, and it ended with a poisoned copy of Anthropic’s Claude Code extension sitting in Cursor’s extension marketplace for eleven hours.
The chain, step by step
Cursor’s investigation pieces the incident together in five stages:
- It started at Fireworks. An attacker obtained an access key Cursor had issued to Fireworks AI, one of its inference and training vendors. The key was, in Cursor’s words, “exposed by Fireworks,” and the attacker used it to copy data from a read-only storage bucket the two companies used to share data. The bucket held limited coding-agent traces — mostly from 2025, from sessions where users had not enabled Privacy Mode and which Cursor was permitted to use for model training. A small number of those traces contained credentials that had not been redacted; fewer than 0.01% of customers had API credentials exposed.
- The credentials opened doors. Starting from the harvested credentials, the attacker reached select SaaS tools Cursor uses, and from there accessed archived non-production source repositories, limited customer support data and email conversations, product usage metrics and contact details, summaries of some coding-agent conversations, and credentials users had sent to the support team.
- The attacker published a malicious extension. On October 29, using a stolen credential from the exposed internal data, the attacker placed modified copies of Anthropic’s Claude Code extension —
Anthropic.claude-codeversions 2.1.312 and 2.1.313, for Windows and macOS — directly into the cache that serves Cursor’s VS Code extension marketplace. At most 5,000 users who requested those cached versions on October 29 potentially received the malicious build. It was available for approximately 11 hours, from 2:27 a.m. to 1:35 p.m. Pacific time, before Cursor removed it, took the marketplace offline, and worked with the cloud provider hosting the attacker’s server to take that down too. - Notification. On October 30, Cursor published a security notice with indicators of compromise, remediation steps, and scripts to check affected systems, and began notifying users directly. Team admins whose credentials were exposed received the specific keys to rotate.
- Containment. Cursor says the last attacker activity it has seen was on November 1, and it was blocked. It has moved all inference traffic off Fireworks, deactivated the key it had issued to the vendor, rotated credentials across its internal systems, redeployed production infrastructure as a precaution, and engaged external forensics experts.
What the malicious extension did
The tampered builds were not a subtle data theft. According to Cursor’s October 29 notice and Microsoft’s independent analysis, injected JavaScript ran when the extension module loaded, before normal initialization. It tried to add an attacker-controlled SSH public key, labelled svc-telemetry, to the user’s authorized keys, enable the local SSH service, and open a reverse SSH tunnel — with persistence registered as a scheduled task on Windows and a LaunchAgent on macOS, so the tunnel would retry in the background. Microsoft’s researchers found the reconnaissance callback collected selected variable names, not their values, and that successful remote access depended on the machine’s permissions and SSH configuration — but malicious files and authorization keys could remain even where the connection failed.
Cursor’s advice to anyone who installed or updated to 2.1.312 or 2.1.313 from its marketplace during the window is blunt: uninstalling the extension alone may not be enough. Its notice lays out the full cleanup steps.
What was not affected
The boundaries Cursor draws matter as much as the breach itself. Privacy Mode users’ prompts and code were not affected; coding-agent traces were exposed for fewer than 0.05% of users, and Cursor says it notified each of them directly — if you received no notification, your data was not in the traces. Customer code repositories, production systems, the Cursor app and its official updates, and release signing keys were not touched; Cursor ships from a separate code hosting system the attacker never accessed. Active inference requests were not exposed. And no other extension in the marketplace was affected — Cursor says it validated that all others were authentic and rebuilt the marketplace store.
Just as important is what the incident was not. Anthropic did not publish the malicious copies, and this was not a vulnerability in Anthropic’s product. The marketplace source, Open VSX, never contained the malware, and the packages were not distributed through Microsoft’s Visual Studio Marketplace or the @anthropic-ai/claude-code npm package. The poison entered through Cursor’s own distribution cache — a cache only Cursor operated.
The attribution question
One point of attribution deserves care. That the key was “exposed by Fireworks” is Cursor’s account. Fireworks’ own security bulletins page, last updated November 9, does not mention Cursor, and it is not public which of Fireworks’ two October intrusions — one detected October 6, another on October 23–24 via a third-party cloud provider — exposed the key. Fireworks has said the evidence links its incident to a financially motivated group that sells stolen LLM access to illicit routing gateways, a crime sometimes called “LLMjacking.” The full picture of the vendor side is still incomplete, and Cursor’s account is, for now, the detailed one.
The blast radius may also be wider than one company. Apify and Epoch AI published their own incident reports on the same day, November 13 — a sign that the Fireworks breach’s consequences are still being mapped across the AI ecosystem.
Why this one matters
The Cursor incident compresses three trends into a single week. First, coding-agent traces have become a secret store: the logs AI tools keep to train better models are dense with credentials, and a single unredacted trace can be the key to an internal system. Second, the AI supply chain is now a chain of vendors — an inference provider’s storage bucket, a marketplace cache, an extension distribution path — and each link trusts the one before it. Third, developer tools are high-value targets precisely because developers’ machines hold the keys to everything else: source code, cloud accounts, production access.
None of that is new in principle. What is new is the scale at which AI coding tools concentrate it. Millions of developers now run agents that read their code, keep transcripts of their sessions, and update themselves from marketplaces — and the companies behind them share those transcripts with vendors under keys that can be copied. Cursor’s response has been thorough and unusually transparent: it named the vendor, published the timeline, quantified the exposure, and told affected users exactly what to rotate. The open question is whether the rest of the ecosystem — vendors included — will be as forthcoming about their own links in the chain.
What to watch
Three things will tell us how far this incident reaches. First, Fireworks’ own account: whether it confirms the key exposure and identifies which intrusion caused it. Second, whether other Fireworks customers publish incident reports of their own in the coming days, as Apify and Epoch AI already have. And third, whether the industry treats agent traces as the sensitive data they evidently are — redacted by default, retained briefly, and never shared with a vendor under a key that can read a whole bucket.








