A vendor key with read-only access can still expose credentials that allow changes somewhere else. In its 13 November investigation update, Cursor says an attacker used a key exposed by Fireworks AI to copy historical coding-agent traces from a shared storage bucket. Some traces contained unredacted credentials. Those credentials opened access to other systems, including support material that held secrets users had sent to Cursor.
The incident began before this disclosure: the tampered extension was distributed on 29 October. The newly published explanation matters to anyone who has pasted a token into an agent conversation or a support ticket. That old conversation can remain a usable credential store long after the immediate task is over.
What Cursor confirms
Cursor says the copied traces came from sessions outside Privacy Mode, mostly from 2025. It reports API-credential exposure for fewer than 0.01% of customers. The attacker reached selected SaaS tools and later used a stolen internal credential to place malicious copies of the Claude Code extension in Cursor's marketplace cache. These findings come from Cursor's investigation; the public record does not independently establish every step.
The Hacker News covered the chain on 15 November. A separate Cloud Security Alliance analysis on 16 November examines the resulting trust boundaries and treats retained agent traces as a secondary store of credentials. Both provide useful independent interpretation of the disclosures, rather than another set of forensic observations.
Keep the extension warning within its actual scope
Cursor's original security notice concerns tampered Anthropic.claude-code versions 2.1.312 and 2.1.313served through its marketplace on 29 October, for Windows and macOS. Cursor says the copies attempted to establish persistent SSH access. Anthropic did not publish these builds. Check the installation source and exposure window against the official notice instead of assuming every Claude Code installation is affected.
For affected installations, Cursor advises investigation of the host, removal of persistence and rotation of accessible credentials and sessions from a trusted system. Uninstalling the extension alone may leave other components behind. Follow the current vendor remediation guidancewith your security team before giving the device replacement access.
What the public timeline still leaves open
Fireworks' 17 November bulletin now describes two separate security incidents. It says both are contained and remediated, but does not identify which one exposed the key Cursor had issued. The time between the theft of that key and the marketplace tampering is also not publicly established. A reconstruction should preserve those gaps.
Cursor reports that the last attacker activity it observed was blocked on 1 November. It also reports no impact to Privacy Mode prompts and code, customer code repositories or its official application releases. Those are the company's findings for this incident; they do not describe what every other AI product retains or prove the safety of a particular developer's machine.
Find the copies before closing the rotation task
Give each exposed credential an owner and an accepting service. Then ask where the value was copied: an agent session, a diagnostic bundle, a ticket attachment, a chat transcript, a command's output or an old configuration export. Record locations and identifiers in the incident ledger, leaving the values themselves out.
Prioritise credentials that still work, especially those with broad access or no expiry. An old trace does not make the credential inside it obsolete. Conversely, finding a retired token does not prove current access. Check its status with the service owner, and distinguish confirmed exposure from the systems it might have reached.
- Rotate where the credential is accepted. Fireworks' guidance explicitly says third-party credentials require action with their issuing providers. Deleting the value from a trace, ticket or vendor settings page does not invalidate it.
- Update the approved consumers. Identify which integrations and jobs use the credential. Install the replacement through the normal runtime secrets store, then test legitimate access and rejection of the retired value.
- Review active access. Use the service's supported session and token controls. Ask the owner what a password or API-key change actually invalidates, and investigate additional access created during the exposure.
- Reduce retained copies. After preserving necessary incident evidence, remove unnecessary sensitive attachments and exports through the relevant retention process. Ask vendors which historical datasets and subprocessors still hold session material.
Make the next support handover smaller
A support team often needs an error code, a request identifier and the name of a failing integration. Start with that information. Provide a redacted example when it answers the question, rather than a complete environment file or a screenshot containing credentials. Verify a request for access through an established support channel, especially during a public security incident.
If someone genuinely needs temporary access, create a narrowly scoped credential for the task, agree who will revoke it and set an end time where the provider supports one. Deliver the value only to the verified recipient. Keep the ticket useful for coordination by recording the owner, transfer status and revocation result instead of the secret itself.
Where Secretus fits
Secretus can transfer a text secret between people with browser-side encryption and a link that becomes invalid after expiry or first successful access. Use trusted devices and a verified recipient for the handover. The credential still needs its own expiry or revocation at the accepting service. A one-time link cannot erase copies already retained in agent traces, clean an affected endpoint or revoke a provider's access key.
Sources
- Cursor: investigation update, 13 November 2026
- Cursor: affected extension versions and remediation, 29 October 2026, updated 30 October
- Fireworks AI: security bulletins, including the 17 November 2026 update
- The Hacker News: Cursor and Fireworks incident reporting, 15 November 2026
- Cloud Security Alliance: independent analysis of the Fireworks–Cursor chain, 16 November 2026
