Cursor breach traced to a key exposed at Fireworks and to coding-agent traces that contained credentials
In its post-mortem published on November 13, Cursor explained that the attack leading to the fake Claude Code extension on October 29 began with an access key Cursor had entrusted to Fireworks, its inference and training provider. Coding agent traces from 2025 contained unmasked credentials, which paved the way into its internal tools. Fewer than 0.01% of customers had API credentials exposed.

On October 30, we reported how a fake version of the Claude Code extension ended up in Cursor’s extension marketplace, carrying an SSH backdoor. One question remained unanswered: how could a package have impersonated Anthropic? Cursor answered it in a post-mortem published on November 13, following a two-week investigation. The vulnerability was neither at Anthropic nor in the marketplace’s upstream source, but at a third-party vendor, and the attack’s fuel was a rarely discussed type of data: coding agent traces.
The Attack Chain, Step by Step
- A key at Fireworks. Cursor had issued an access key to Fireworks, its inference and training provider, to share data in a read-only storage space. According to Cursor, Fireworks exposed this key, and the attacker used it to copy the contents of that storage space.
- Agent traces with secrets. This space contained coding agent traces, primarily from 2025, originating from sessions without Privacy Mode that Cursor was authorized to use for training its models. A portion contained unmasked credentials: keys that users had let slip into their sessions. Fewer than 0.01% of customers had API credentials exposed.
- Pivoting to internal tools. Using these credentials, the attacker reached certain Cursor SaaS tools: archived non-production code repositories, support data and emails, usage metrics, agent conversation summaries, and even credentials that users had submitted to support.
- The backdoored extension. On October 29, using a credential stolen from these systems, the attacker placed backdoored copies of the Claude Code extension (versions 2.1.312 and 2.1.313) directly into the cache serving Cursor’s marketplace. Open VSX, the marketplace source, never contained the malware. At most 5,000 users may have received it, over a window of approximately 11 hours.
Cursor says it observed the attacker’s last activity on November 1 and blocked it. The application, its official updates (distributed from a separate system, with no signing key exposed), production infrastructure, and customer repositories were not impacted, according to the developer. Affected users were notified individually: without a notification, your data was not among the exposed traces.
What Fireworks Says
Fireworks published its own update on November 9. A third-party cloud provider hosting its research machines was compromised on October 23; the following day, the intruder pivoted to Fireworks’ development machines. The company observed access to billing information there, and for certain customers, access to secrets, datasets, and weights of models undergoing training. It links the attack to a cybercrime group specializing in “LLMJacking”: stealing access to language models to resell tokens through illicit gateways. Its bulletin does not mention Cursor or the key in question: the link between the two incidents rests on Cursor’s account.
What Cursor Changed
Cursor migrated all its inference traffic off Fireworks, revoked the key, and cut the vendor’s access to all shared data. It rotated all internal credentials, even absent evidence of exposure, restricted the network usage of credentials at its vendors, added secret detection across its repositories and other internal data sources, and redeployed all of its production.
Our take
This is the third incident in a month where AI-related secrets have been targeted: Railway saw AI vendor keys stolen from its customers, Fireworks describes attackers reselling model access, and Cursor shows how agent session traces served as an entry point. There is no evidence suggesting the same group is responsible. The practical lesson, however, is shared: a coding agent session records everything passing through the terminal, including keys pasted “just to test.” Once these traces are used for training and shared with a vendor, they turn into a pool of secrets. For a team, this means enabling Privacy Mode whenever possible, never pasting plain-text secrets into an agent or support ticket, and promptly rotating any key that has touched a session. For developers, the bare minimum is redacting secrets before storing or sharing a trace, which was not done for a portion of the traces Cursor shared.
Sources
- Cursor — An update on our recent security incident (November 13, 2026) Primary source
- Cursor (X) — incident post-mortem announcement Primary source
- Fireworks AI — Update on October 2026 security incident (November 9, 2026) Primary source
- Cursor — Security notice: malicious version of Anthropic.claude-code (2.1.313) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.


