Ground Truth.
AI, checked against the source.

News · 2026-11-15

Cursor links malicious extension to exposed coding-agent traces

Cursor says an attacker used credentials found in exposed historical coding-agent traces to reach selected company services and place malicious Claude Code-branded extensions in its marketplace cache. The November 13 update links an October software-supply-chain attack to an earlier exposure involving AI infrastructure provider Fireworks.

Key facts

A coding assistant’s history can contain much more than prose. A trace can include snippets of source code, tool outputs and, if a secret was pasted or printed, credentials. Cursor says a limited number of older traces in its shared bucket contained unredacted credentials. The company attributes exposure of its access key to Fireworks and says credentials in the copied data subsequently enabled access to selected software-as-a-service tools and the marketplace cache.

The chain resembles a set of keys left in successive rooms. A read-only key might not itself alter software, but a document in the room can contain a different key with different permissions. Cursor has not publicly identified the particular credential or service that permitted the cache write, so a more specific route would be speculation.

Fireworks’ incident update confirms unauthorized access and customer-data impact in October, but does not name Cursor’s key or the bucket. Cursor’s account is therefore the source for the specific cross-company connection. The distinction matters: compatible timelines are not independent confirmation of every link in the chain.

The affected files bore the name of a familiar coding extension. Microsoft Security Research’s analysis says the inserted code attempted to add an attacker’s SSH key, establish a reverse connection and create persistence. Its examination distinguishes attempted actions from a working remote login: success depended on the machine’s setup and privileges. Cursor’s original security notice gives affected versions, indicators and remediation instructions, including why removing the extension alone may not remove persistence.

Cursor says fewer than 0.05% of users had non-Privacy-Mode traces exposed and less than 0.01% of customers had API credentials in those traces. Those are different measures, not estimates of how many machines were compromised. Cursor says it notified affected users, disabled the key and rebuilt the marketplace. The company’s conclusions about unaffected systems and the scope of customer impact remain its investigation findings rather than an independently published full forensic audit.

The attack illustrates why AI tool supply chains need controls around the whole delivery path, including mirrors and caches, not only the original publisher. A verified upstream package is useful only if the copy ultimately served to a user remains authentic. Similarly, retention, access restrictions and secret scanning matter for traces sent to vendors, even when their immediate purpose is model development.

Cursor labels its bucket “read-only”; that constraint did not prevent an attacker from copying data and exploiting other secrets contained in it. The evidence does not establish that artificial intelligence conducted this intrusion, or that every potentially exposed user installed or ran the payload. The actionable conclusion is narrower and more useful: records created by AI-assisted development became part of an ordinary credential and software-distribution attack.

The data exposure and extension distribution are separate harms. A user could be affected by a leaked trace without installing a malicious extension; another could have requested the malicious extension without appearing in the exposed trace subset. Cursor says the traces were mostly from 2025 and came from sessions outside Privacy Mode that were eligible for model training. It also says some limited support and business data were reached through the subsequent SaaS access. These are the company’s investigation findings, and they should not be generalized to all users.

The marketplace was a particularly consequential handoff because developers rely on extension updates to deliver executable code. A recognizable package name can help someone find a legitimate tool, but its name alone does not prove the bytes came from the original author. The malicious builds were inserted into the cache serving Cursor’s marketplace, according to Cursor, rather than being a compromise of the official upstream package. That distinction narrows both the affected download path and the investigative question.

For a team responding to a possible installation, the version and time window matter. Cursor’s notice provides indicators and platform-specific steps. Microsoft notes that artifacts may remain even when a reverse connection did not succeed. An organization cannot infer a successful compromise merely from eligibility to receive the package, but neither should it assume that uninstalling an extension is sufficient evidence that every persistence mechanism was removed.


Primary source, verified: read the paper →

Key questions

Did 5,000 people run the malicious extension?

That is not established. Cursor says at most 5,000 users who requested the affected cached versions could have received them.

Was Anthropic’s official extension compromised?

Microsoft says the malicious packages were not distributed through Anthropic’s npm package or Microsoft’s marketplace; Cursor says the upstream Open VSX source was clean.

What did the malicious code attempt?

Microsoft says it attempted to install an SSH key, establish a reverse tunnel and create persistence, subject to host permissions and configuration.
Cite this

APA

Ground Truth. (2026, November 15). Cursor links malicious extension to exposed coding-agent traces. Ground Truth. https://groundtruth.day/news/cursor-traces-fireworks-malicious-extension-update.html

BibTeX

@misc{groundtruth:cursor-traces-fireworks-malicious-extension-update,
  title  = {Cursor links malicious extension to exposed coding-agent traces},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/cursor-traces-fireworks-malicious-extension-update.html}
}

Topics: cybersecurity · ai-security · supply-chain · credentials · coding-agents

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.