Trojanized Claude Code Builds Sat in Cursor's Marketplace for 11 Hours
Key Facts
- Malicious Claude Code builds — versions 2.1.312 and 2.1.313 for macOS and Windows — were available in Cursor's extension marketplace for approximately 11 hours before removal.
- The extension sent host details and backdoor status to the attacker's server at the IP address 134.209.82[.]225.
- According to Phorion, coding errors in the malware's SSH setup routines likely prevented attackers from gaining remote access to many affected Windows and macOS systems.
- Fewer than 0.01% of Cursor customers had API credentials exposed after some traces included unredacted credentials.
- Cursor's estimate that the incident could have affected as many as 5,000 users is a maximum figure, not a confirmed number of compromises.
A short window and two specific builds
Attackers distributed malicious versions of Anthropic’s Claude Code extension through Cursor’s extension marketplace. Cursor’s scope statement names two builds: versions 2.1.312 and 2.1.313, for both macOS and Windows. They were available for approximately 11 hours before Cursor removed them.
The condition for exposure is narrow, and it is worth reading twice. It applies to anyone who installed or updated Claude Code to 2.1.312 or 2.1.313 from Cursor’s marketplace on macOS or Windows, with the timeline beginning at 2:27 a.m. A developer who kept an older version and did not update sits outside that window. So does anyone who installed Claude Code from another source.
Eleven hours is a working day in one time zone and the middle of the night in another. That is the sort of detail that decides who was actually at a keyboard while the malicious builds were live.
What the builds did
Once running, the malicious extension sent host details and backdoor status to the attacker’s server at the IP address 134.209.82[.]225. That is a check-in rather than a full intrusion: it tells whoever is listening which machines are reachable and whether a backdoor is in place.
According to Phorion, the second part of that plan largely failed. Coding errors in the malware’s SSH setup routines likely prevented attackers from gaining remote access to many affected Windows and macOS systems. A backdoor that cannot finish its own setup is still evidence of intent, and still a machine that needs to be rebuilt. It is not the same thing as an attacker holding a shell.
That distinction shapes how the incident should be read. The most alarming version of the story, remote control of developer laptops for eleven hours, is not what the evidence supports so far.
The credentials were the damage that traveled
Some traces included unredacted credentials. As a result, fewer than 0.01% of Cursor customers had API credentials exposed.
Small percentages and developer tooling are a bad combination. A fraction of a percent of a large developer population is still a set of live keys, and live keys do not need a working backdoor to be useful. They can be replayed from anywhere.
Cursor reports that the incident could have affected as many as 5,000 users, but that figure is a maximum estimate, not a confirmed number of compromises. The company draws that line itself. Treat 5,000 as a ceiling, and treat “fewer than 0.01%” as the number describing confirmed credential exposure. The two are not interchangeable, and coverage that blurs them will overstate what is known.
The trail also runs past Cursor’s marketplace. The draft coverage behind this analysis describes Cursor and Notion disclosing how the breach affected them over the following weeks, alongside a separate Fireworks AI disclosure involving exposed API keys and tokens and another incident in which attackers reached development machines through a compromised third-party cloud provider. Those events belong to the same story about credentials and third parties. They do not belong to Cursor’s scope statement, and figures from one should not be used to describe the other.
What the timeline says about marketplace trust
The mechanics here are ordinary, which is the uncomfortable part. An extension is published, accepted by developers who trust the registry that served it, and removed once someone notices. Eleven hours is quick by the standards of a decade ago and slow by any standard that matters to the person who updated at 2:30 a.m.
Cursor has published both an incident update and a dedicated security notice naming the affected versions, and those primary documents are worth reading in full rather than in summary. The useful questions for anyone installing from a marketplace are dull ones. Which version is installed? Where did it come from? When was it last updated? An extension inventory that answers those three questions turns an eleven-hour window into a five-minute check.
The lesson is not that curated marketplaces are worthless. It is that the gap between publication and removal is measured in hours, and the recovery work lands on the people who installed inside that gap. Rotate the credentials, check the update history, and write the version numbers down somewhere other than memory.
Sources & References
- An update on our recent security incident · Cursor. Primary official statement. https://cursor.com/blog/security-incident-update
- Artificial Intelligence - GitGuardian Blog - Take Control of Your Secrets Security. Primary research blog. https://blog.gitguardian.com/tag/artificial-intelligence/
- Secrets detection - GitGuardian Blog - Take Control of Your Secrets Security. Primary research blog. https://blog.gitguardian.com/tag/secrets-detection/
- Supply Chain Security - GitGuardian Blog - Take Control of Your Secrets Security. Primary research blog. https://blog.gitguardian.com/tag/supply-chain-security/
- Anna Dovbysh - GitGuardian Blog - Take Control of Your Secrets Security. Primary research blog. https://blog.gitguardian.com/author/anna-dovbysh/
SecurityXP delivers cybersecurity news, vulnerability analysis, data breach reports, and threat intelligence.
Security Digest
The most important cybersecurity stories, CVE alerts, and threat intelligence — one concise email. Unsubscribe anytime.
Related Articles
Microsoft restricts employee Claude Fable 5 access over Anthropic data retention
Microsoft restricts employee access to Claude Fable 5 while legal reviews Anthropic's 30-day retention policy, which can retain flagged content for two years.
Feds & npm Issue Supply Chain Security Guidance
Faster development times, innovation, and a thriving open-source community have all been made possible by the ability to use another developer's project as a dependency. With many JavaScript projects
Auditing the State Transfers That Execution Traces Miss in Long-Horizon LLM Agents
A paper submitted on 17 Nov 2026 to Computer Science Cryptography and Security, “Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing…
FlowSH Correlates Anomalous NetFlows With Server-Side Logs to Catch SQLi That ModSecurity Misses
FlowSH, a detection system posted 18 November 2026 under Computer Science Cryptography and Security, combines unsupervised machine learning on network…