Cursor disclosed on November 13, 2026 that an access key exposed at Fireworks led an attacker from stored coding-agent traces into internal tools and, ultimately, its extension marketplace cache. The forensic account explains the October 29 malicious Claude Code extension incident. For university research software teams, it connects retained AI-session data with the integrity of developer tools. Cursor published the findings after a two-week investigation.
From a vendor key to the marketplace cache
According to Cursor, the attacker obtained a key it had issued to Fireworks and copied information from a read-only storage bucket shared with that vendor. The bucket held limited coding-agent traces, mostly from 2025, from sessions outside Privacy Mode that Cursor says it was permitted to use for training. Some traces retained credentials that had escaped redaction. The Hacker News described that sequence in its November 15 report.
Those credentials opened a path into selected software-as-a-service tools used by Cursor. The company says the material reached through them included archived non-production source repositories, some containing business contacts; limited support records and email conversations; product-usage metrics and contact information; summaries of some agent conversations; and credentials users had sent to support. These are Cursor’s investigation findings, rather than evidence that every customer or repository was exposed. Its account distinguishes these internal records from customer code repositories.
On October 29, a credential obtained from exposed internal data let the attacker put unauthorized extension copies directly into the cache serving Cursor’s VS Code extension marketplace. Cursor says Anthropic did not publish them, the incident was not a vulnerability in Anthropic’s product, and Open VSX, the upstream marketplace source, never held the malware. The distribution cache was therefore a separate point of compromise from the original extension publisher. BleepingComputer reported the newly disclosed chain on November 14.
Three exposure figures describe different groups
Cursor’s November 13 statement gives the following measures. They describe different forms of exposure and use different denominators; they cannot be combined into one compromise rate.
| Measure | Cursor’s reported scope | What the number describes |
|---|---|---|
| API credentials exposed | Less than 0.01% of customers | Customers whose API credentials appeared in exposed traces. |
| Coding-agent traces exposed | Fewer than 0.05% of users | The trace exposure described for users outside Privacy Mode. |
| Potential malicious-build delivery | At most 5,000 users | Users requesting specific cached extension versions on October 29 who potentially received the build. |
The last figure is an upper bound on potential delivery, not a count of successful backdoor installations. Cursor’s original security notice identifies versions 2.1.312 and 2.1.313 for macOS and Windows, available on October 29 from about 2:27 a.m. to 1:35 p.m. Pacific time. The November account describes an approximately 11-hour distribution window.
The notice says the modified extension collected host information and attempted to establish a persistent SSH backdoor. Enabling the local SSH service generally required administrator privileges the extension lacked, so that step failed on most hosts, according to Cursor. However, the company says an unauthorized key and persistence mechanisms could remain. This is why a failed remote-login attempt does not establish that a machine was unaffected, and why Cursor warns that uninstalling the extension alone may be insufficient.
What the November disclosure adds
CASRAI’s earlier report, “Fireworks AI widens breach disclosure to datasets and model weights”, covered Fireworks’ November 9 bulletin and explicitly noted the absence of a detailed public Cursor statement. The November 13 account supplies that missing vendor-to-Cursor sequence. It is a forensic disclosure about the October incident, rather than evidence of a new attack on November 13.
In its November 9 bulletin, Fireworks separately described an October 23 incident at a cloud provider hosting its research-and-development machines, followed by unauthorized access to Fireworks development machines on October 24. Fireworks said it had detected no further successful unauthorized access to its customer platform after October 25. That date concerns Fireworks’ environment; Cursor’s later observation concerns its own systems.
Outside reporting also preserves the distinction between publisher and distributor. SecurityWeek’s November 2 coverage quoted Anthropic’s Felix Rieseberg saying Anthropic had neither built nor published the tampered copy and that its official channels were unaffected. The newer reports explain how the attacker reached Cursor’s distribution path; they do not reverse that earlier attribution.
Why the chain matters for research software teams
For research offices, the concrete issue is the connection between an AI supplier’s retained data and a development workstation’s software supply chain. In this account, a read-only bucket still contained credentials useful elsewhere. The institutional implication is that assessing third-party AI vendor risk includes the data exchanged for training and analysis, alongside the service researchers actively use. This is an interpretation of the disclosed chain, not a finding that university systems were breached.
The retention distinction also matters. Cursor says active inference requests and current live traffic were not exposed, while the copied traces came from historical offline-analysis storage. Fireworks says it has detected no impact to its Inference, Training SDK and Nexus products, which it describes as zero data retention by default. These statements concern specified data paths and products; they are not interchangeable assurances about all stored information. CASRAI’s background on zero data retention for AI tools in research explains why scope matters.
Cursor says it contacted affected customers and sent team administrators lists of specific exposed keys to rotate, while asking issuing services to revoke those credentials. For an institutional data breach response plan, that separates a vendor notification about stored data from endpoint evidence about an installed extension. Neither the percentage figures nor the delivery ceiling identifies an individual laboratory’s exposure.
Containment and the remaining response
Cursor says it believes the incident is contained: the last attacker activity it observed was on November 1 and was blocked. It reports moving inference traffic away from Fireworks, rotating credentials, rebuilding the extension marketplace, expanding secret scanning and engaging external forensic experts. Its exclusions include Privacy Mode prompts and code, customer repositories, the production systems running Cursor, and official application updates. These remain the company’s stated investigation conclusions. Cursor says it will contact customers if further findings affect them.
For the affected extension versions, Cursor continues to direct customers to its security notice with indicators and remediation scripts. Fireworks says its investigation remains ongoing and that self-serve dataset uploads and downloads for Batch Inference and dataset-based training were disabled during the response. Neither statement sets a dated completion milestone; both commit to further updates as appropriate.
Sources
- Cursor, “An update on our recent security incident”, November 13, 2026. Checked 2026-11-16.
- Cursor, security notice on malicious Anthropic.claude-code versions, dated October 29, 2026; updated October 30, 2026. Checked 2026-11-16.
- Fireworks AI, “Update on October 2026 security incident”, updated November 9, 2026. Checked 2026-11-16.
- BleepingComputer, Lawrence Abrams, “Cursor confirms security breach led to malicious Claude Code extension”, November 14, 2026. Checked 2026-11-16.
- The Hacker News, Ravie Lakshmanan, “Cursor Says Stolen Fireworks AI Key Led to Trojanized Claude Code Extension Incident”, November 15, 2026. Checked 2026-11-16.
- SecurityWeek, Ionut Arghire, “Malicious Claude Code Extension Served in Cursor Supply Chain Attack”, November 2, 2026. Checked 2026-11-16.







