For about 11 hours on October 29, developers who installed or updated Anthropic’s Claude Code extension inside Cursor may have received a version that tried to open a secret door into their computer. Cursor has now explained how that happened, and the answer is more unsettling than a single bad upload. It started with one stolen key at a different company entirely.
In an incident report published Friday, Cursor said attackers used an access key stored at Fireworks AI, the inference provider it relied on, to copy old coding agent logs. Those logs contained passwords and API keys that nobody had scrubbed. The attackers followed that trail into Cursor’s internal tools and eventually into the cache that serves its extension marketplace, where they planted two trojanized builds of Claude Code.
The short version
- Malicious versions: Anthropic.claude-code 2.1.312 and 2.1.313, served only through Cursor’s marketplace cache
- Exposure window: October 29, about 2:27 a.m. to 1:35 p.m. PT
- Up to 5,000 developers may have received a bad build
- The payload tried to install an SSH backdoor with a key labeled
svc-telemetry - Anthropic’s own extension, npm, Open VSX and Microsoft’s marketplace were not compromised
How one key became a supply chain attack
The most important lesson in Cursor’s report is that a “read only” credential is not harmless if the data it can read contains other credentials. Here is the chain as Cursor describes it.
Cursor says the exposed traces affected fewer than 0.05% of its users, and fewer than 0.01% of customers had API credentials in them. Those are small percentages, but they were enough. The attackers also took archived non production source code, some support emails, product usage metrics, summaries of some agent conversations, and credentials that customers had previously pasted into support tickets.
Fireworks has had a rough autumn. It disclosed a breach in early October involving exposed API keys and tokens, then a second one after attackers got into its development machines through a compromised third party cloud provider on October 23 and 24. Notion has since said AI chat transcripts from about 1% of its customers were exposed through the same vendor. Fireworks has described the attackers as financially motivated and focused on “LLMjacking,” which means stealing access to AI models and reselling it.
What the malicious extension tried to do
Researchers at Phorion and OpenSourceMalware found that the attackers added a JavaScript loader to the top of the extension’s extension.js file and left everything else alone. Claude Code kept working normally, so nothing looked wrong. The hidden code ran as soon as the extension loaded, without the developer touching Claude at all.
- Collected the username, hostname, public IP address, Git email and the names of environment variables that looked like secrets
- Added an attacker controlled SSH public key with the comment
svc-telemetry - Tried to enable the local SSH server and open a reverse tunnel to
134[.]209[.]82[.]225 - Set up persistence through a macOS LaunchAgent or a Windows scheduled task and Startup script that retried roughly every minute
There is some good news. The malware was sloppy. On macOS it checked whether Remote Login was enabled without admin rights, got an error instead of “Off,” and so never turned SSH on. On Windows it tried to install OpenSSH by calling /bin/sh, a Unix shell that does not exist on a normal Windows machine. Cursor says those mistakes and permission requirements blocked remote access on most systems.
That does not mean you are in the clear. Even where the tunnel failed, the malware could still write its SSH key and startup files to disk and send your machine details home. And if SSH was already running on your Mac or PC, which is common for developers who use remote workflows, the backdoor had a real chance of working.
Were you exposed? Quick reference
| Factor | Higher risk | Lower risk |
|---|---|---|
| Editor | Cursor | VS Code, JetBrains, terminal only |
| Install source | Cursor’s built in extension marketplace | npm package @anthropic-ai/claude-code |
| Timing | Installed or auto updated on Oct 29, 2:27 a.m. to 1:35 p.m. PT | Last update before Oct 29 or after the fix |
| Version seen | 2.1.312 or 2.1.313 | Any other version |
| SSH server | Already enabled (Remote Login on, OpenSSH running) | Never enabled |
Auto updates matter here. Being on a newer version today does not prove an older bad build never ran, so do the checks below even if your extension looks current.
How to check your machine in two minutes
These are the indicators published by Cursor and Microsoft’s security researchers. A hit on any of them means you should treat the machine as compromised.
macOS (Terminal):
# Look for the attacker's SSH key
grep -n "svc-telemetry" ~/.ssh/authorized_keys
# Check the key fingerprint
ssh-keygen -lf ~/.ssh/authorized_keys | grep "0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk"
# Look for the persistence files
ls -la ~/Library/LaunchAgents/com.apple.sysmond.plist ~/.config/.smt ~/.config/.smtkWindows (PowerShell):
# Look for the attacker's SSH key (user and admin files)
Select-String -Path "$env:USERPROFILE\.ssh\authorized_keys" -Pattern "svc-telemetry"
Select-String -Path "C:\ProgramData\ssh\administrators_authorized_keys" -Pattern "svc-telemetry"
# Scheduled task and Startup fallback
Get-ScheduledTask -TaskName "SysMon" -ErrorAction SilentlyContinue
Test-Path "$env:APPDATA\Microsoft\Windows\Start Menu\Programs\Startup\sysmon.cmd"
# Dropped script and tunnel identity
Test-Path "$env:USERPROFILE\.config\smt.cmd"
Test-Path "$env:USERPROFILE\.config\smtk"Careful with “SysMon”: that name is generic, and Microsoft’s own Sysinternals tool is called Sysmon. Do not delete a task just because of its name. Check what it runs. The malicious one points at the smt.cmd script. Also, a clean result does not prove the code never ran, so if you know you installed a bad version, keep going with the steps below.
If you find something
- Disconnect the machine from the network and, if you work for a company, tell your security team before deleting anything. They will want the files as evidence.
- Remove only the bad key. Delete the single
svc-telemetryline fromauthorized_keysand keep your legitimate keys. - Remove persistence: the LaunchAgent and
.smtfiles on a Mac, or the scheduled task, Startup script and.configfiles on Windows. Cursor offers official cleanup scripts for both platforms. Read them before you run them. - Rotate every secret that lived in that machine’s environment variables, from a clean device. Think cloud keys, GitHub tokens and AI API keys.
- Rebuild if in doubt. If you cannot be sure what happened after the extension loaded, a clean reinstall is the only real fix.
The bigger problem: AI logs are full of secrets
This is not just a Cursor story. Security firm Wiz has nicknamed this wave of incidents “Black October,” and it has hit Railway, Northflank, Lightning AI, OpenCode, Warp and others. At Railway, stolen customer secrets were used to run up around $740,000 in fraudulent AI charges, which is why we told readers to rotate their secrets after the Railway breach last month.
The pattern is the same each time. Coding agents see everything a developer sees, including the .env file, the pasted token and the database URL. When those sessions are logged and shipped to a third party for training or debugging, the logs become a treasure chest. It is the same risk that pushed Microsoft to start treating AI agents like insider threats: an agent with broad access is a powerful tool and an equally powerful leak.
Cursor says it has moved all inference traffic off Fireworks, rotated credentials, rebuilt its marketplace and redeployed production. It also says Privacy Mode users’ code and prompts were not in the exposed data. If you use any AI coding tool, that last point is the practical takeaway. Turn on the strictest privacy setting your tool offers, keep secrets out of prompts and chat windows, and assume anything an agent can read might one day end up somewhere you did not expect.
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

