An exposed vendor key opened historical agent traces with credentials. A later credential let attackers serve malicious Claude Code builds for 11 hours; at most 5,000 requesters potentially received one.
The company’s November 13 incident update says an attacker obtained a Cursor-issued access key that Fireworks AI exposed. The key allowed the attacker to copy data from a read-only storage bucket Cursor used to share coding-agent traces with the provider. Some of those traces contained credentials that had not been redacted.
Cursor says credentials recovered from the traces opened selected SaaS tools. Separately, the company says the attacker later used a stolen credential from exposed data in an internal system to place malicious copies of Anthropic’s Claude Code extension directly into the cache serving Cursor’s VS Code extension marketplace.
Cursor’s disclosure joins retention and software distribution in one risk chain: credentials left inside historical AI traces created a route toward systems capable of putting code on developers’ machines.
Cursor says an exposed vendor key opened stored coding-agent traces; a later credential from exposed internal data was used to place unauthorized copies in the marketplace cache.
A read-only bucket held live authority
Cursor says the exposed bucket held a limited set of coding-agent traces, mostly from 2025. The sessions were outside Privacy Mode, and Cursor says it was permitted to use them for model training.
Fewer than 0.05% of users had traces exposed, according to the company. Fewer than 0.01% of customers had API credentials exposed. The first figure counts users with exposed traces; the second counts customers with exposed API credentials. The privilege carried by each surviving credential determined the practical blast radius.
Cursor describes the bucket access as read-only: the attacker copied its contents rather than modifying them. The surviving credentials still carried the permissions granted by the services that trusted them.
Fireworks’ November 9 bulletin says its broader October investigation expanded from stored secrets to account and billing information and, for a limited set of accounts, secrets, datasets and incomplete models. Based on its investigation at that time, Fireworks said it had not detected impact to its zero-data-retention Inference, Training SDK and Nexus services.
Fireworks does not name Cursor in that bulletin or identify which part of its investigation exposed Cursor’s key. Cursor’s report establishes the specific access-key and coding-trace chain.
Cursor says some historical coding-agent traces contained unredacted credentials; fewer than 0.01% of customers had API credentials exposed.
The traces supplied the next rung
Cursor says credentials recovered from the traces gave the attacker access to selected SaaS tools it used internally.
From those tools, the attacker accessed selected archived, non-production Cursor 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 Cursor’s support team.
The incident moved through a sequence of credentials with different reach. The vendor key opened retained data. Secrets inside that data opened selected internal services. Cursor then found that a later stolen credential from exposed internal data reached its marketplace cache; its report does not name the exact internal system that held it.
Cursor’s monitoring identified unauthorized activity. The company deactivated the Fireworks key, began rotating exposed credentials and searched for further access. It also temporarily disabled internal and product services while investigating.
One internal credential reached the cache
On October 29, Cursor says the attacker used a stolen credential from exposed data in its internal system to place unauthorized versions 2.1.312 and 2.1.313 of Anthropic.claude-code directly into the cache serving its marketplace.
Anthropic did not publish the modified packages. Cursor says this was not a vulnerability in Anthropic’s product, and Open VSX did not contain the malware. Cursor also says no other marketplace extension was altered.
The malicious builds were available for approximately 11 hours. Cursor estimates that at most 5,000 users who requested those specific cached versions potentially received one. BleepingComputer’s November 14 report brought the newly disclosed credential chain together with the earlier distribution event.
Cursor says the attacker placed unauthorized Claude Code copies directly into the marketplace cache; Anthropic and Open VSX did not publish them.
Eleven hours left a host-level case
Cursor’s October 29 security notice says the modified extension collected host and user details, added an attacker-controlled SSH key, attempted to enable SSH and open a reverse tunnel, and installed mechanisms intended to keep the tunnel running.
Coding errors and missing privileges meant the SSH-enablement step would fail on most Windows and macOS systems where Cursor was not running with administrator rights. The extension could still add the unauthorized key, install persistence and report host details.
Treat potential receipt, installation, extension load, payload execution, SSH enablement, tunnel establishment and interactive access as separate states. The svc-telemetry key or documented persistence shows that the payload changed the host; host and network telemetry must carry the case further. Cursor says the code reported host details and environment-variable names, not credential values.
Start with the svc-telemetry SSH key comment, extension versions 2.1.312 and 2.1.313, and Cursor’s exposure window: approximately 2:27 a.m. to 1:35 p.m. Pacific on October 29.
Cursor advises teams to search endpoint, network and extension-installation telemetry for its published indicators. A matching key entry establishes impact on that machine. Investigators should then look for the documented persistence files, launch mechanisms, scheduled tasks and reverse-tunnel processes.
Affected hosts should be isolated before removing the extension, key and persistence. Uninstalling the package alone may be insufficient because the payload can leave scripts outside the extension. Teams should decide whether to rebuild the host, then rotate credentials and invalidate sessions reachable from it using a trusted machine.
Cursor warns that uninstalling the extension alone may not remove the unauthorized SSH key and standalone persistence.
Map retention to distribution authority
Cursor says it moved inference traffic away from Fireworks, rotated exposed and internal credentials, rebuilt the marketplace, expanded network restrictions across service providers and redeployed production infrastructure. The company believes the incident is contained and says the last attacker activity it observed was blocked on November 1.
Teams need an inventory that connects every retained data store to the credentials and distribution systems it can expose.
Coding traces retained for training or analysis need secret detection before storage, throughout retention and before transfer to another provider. Vendor credentials should be short-lived, narrowly scoped and restricted to their intended network location. Support systems and archived repositories need the same scanning because they can hold working secrets outside production.
Software-distribution credentials deserve a separate tier of protection. They should never appear in general-purpose SaaS tools; cache or marketplace changes need independent approval, tamper-evident records and an immediate way to invalidate every artifact produced during a suspected compromise.
The path to Cursor’s marketplace ran through historical storage after the vendor key was exposed. Closing this class of incident means treating retained agent traces and software-distribution authority as adjacent parts of the same blast radius.






