Cursor: how archived secrets became a software delivery risk

Contents
At a glance
The Cursor and Fireworks incident reports explain how retained credentials can connect a data leak to extension delivery and endpoint cleanup.
The connection between storage and software delivery
A historical dataset can become a route into a current system when it contains a credential that still works. Cursor’s incident update of 13 November 2026 gives a concrete account of that connection: a vendor access key led to stored coding traces, credentials in those traces opened other services, and a further stolen credential was used to place a malicious extension in a distribution cache. This is the company’s reconstruction of its incident, not an independent forensic finding.
The useful question for organisations is how to stop one permission from turning into another. A storage policy, an extension policy and a workstation response procedure normally belong to different teams. Here their responsibilities meet. Our analysis separates those boundaries so that a decision to close one route is not mistaken for completion of the entire response. [1][2]
Read-only access can expose a second permission
Cursor says the original bucket permitted reading and contained historical traces from sessions outside Privacy Mode. Read-only describes what the storage credential could do to that bucket. It does not describe what an attacker could do with a different credential found inside a file. The distinction matters when an inventory classifies an archive as low risk because nobody can modify it.
Fireworks’ separate bulletin, updated on 9 November, describes unauthorised access involving its development environment and additional customer data. It reports revoking credentials, reducing excessive privileges and restricting network paths. These measures address different parts of an access chain. Revocation invalidates a particular secret; privilege reduction limits what a replacement can do; network restrictions can reduce where it can be used. None is interchangeable with deleting a copy of an old dataset. [1][4]
A distribution cache is part of the trust chain
According to Cursor, the malicious copies were inserted directly into its extension marketplace cache on 29 October. The company says Anthropic did not publish them and the upstream Open VSX marketplace did not contain them. That distinction identifies the delivery boundary which failed. A familiar publisher name is insufficient to explain the provenance of a specific downloaded package.
For procurement and security teams, the analytical consequence is to include intermediate distribution systems in the review. Ask who can replace cached packages, how the package reaching a user is matched to the approved upstream version, and which records identify the package actually installed. These are review questions, not claims that a particular control was absent in Cursor. A vendor’s official app, its release process and the extensions it distributes can have different access paths. [2]
Why stopping distribution does not clean a device
Cursor’s security notice identifies the affected extension versions as Anthropic.claude-code 2.1.312 and 2.1.313 on Windows and macOS in the stated 29 October window. It describes attempts to add an SSH key and establish persistence. It also explains that privileges and errors in the malicious code affected whether particular steps succeeded. Attempted installation, successful execution and subsequent access must therefore remain separate findings.
The notice directs affected customers to inspect indicators, isolate hosts with evidence of execution, remove persistence and rotate accessible credentials from a trusted system. Removing an extension alone may leave components which it installed outside its own directory. This is why a marketplace takedown and a workstation recovery are different jobs: one prevents further delivery, while the other establishes the state of an already exposed machine. [3]
Separate the response into three jobs
The first job is to identify exposed data and the credentials it contained. Owners need to know which system issued each credential and whether it remains valid. The second is to reconstruct subsequent access: service logs, package records and endpoint evidence answer different questions. The third is recovery, including replacement credentials and a documented decision on affected devices. This division is our operational reading of the two vendors’ reports.
Consider a hypothetical team with an archived support conversation containing a service token. Deleting the conversation removes one stored copy but does not invalidate the token. Revoking the token stops its future use but does not undo earlier changes made with it. Rebuilding a machine addresses its software state but does not automatically revoke sessions elsewhere. The example illustrates why the completion evidence should match the particular action. [1][4]
Three boundaries and their completion evidence
| Boundary | Question | Evidence to collect |
|---|---|---|
| Stored secrets | Can the exposed credential still be used? | Revocation or expiry record from its issuer |
| Software delivery | Which package reached the device? | Installation record and verified package identity |
| Endpoint recovery | What persisted after installation? | Device investigation and remediation record |
What organisations should ask their suppliers
A practical supplier review starts with retained material: what is stored for evaluation, support or development, for how long, and under which separate permissions? Ask how secrets are detected before storage and when historical collections are rescanned. Next establish who can revoke credentials issued by another service and how the organisation receives a specific, actionable notification. A generic incident email is less useful than a list of affected access paths and owners.
Finally, require closure evidence for each boundary rather than a single completed checkbox. The documents support an analysis of linked access risks; they do not establish cryptocurrency losses or a compromise of every user of either provider. For teams handling exchange integrations or financial applications, the relevance is the access model: credentials retained in working material should be reviewed according to the authority they grant today, even when the material itself is old. [1][4]
Evidence & sources
Source record: · Checked:
- Source reported
Cursor describes access to historical traces through a read-only bucket.
Supporting excerpt: Cursor incident reconstruction
read-only storage bucket
Cursor incident reconstruction ↗ - Source reported
Cursor says malicious extension copies were placed in its distribution cache.
Supporting excerpt: Cursor incident reconstruction
directly on the cache
Cursor incident reconstruction ↗ - Source reported
Cursor identifies the affected extension versions and platforms.
Supporting excerpt: Cursor malicious extension notice
2.1.312 and 2.1.313 for Windows and MacOS
Cursor malicious extension notice ↗ - Source reported
Fireworks describes credential revocation and replacement.
Supporting excerpt: Fireworks incident update
revoked and replaced credentials
Fireworks incident update ↗


