Actively exploited Security news
Cursor traces its breach to a key issued to Fireworks AI, which led to stolen data and a tampered Claude Code extension
The Cursor code editor published the findings of a two-week investigation on 13 November 2026. An attacker obtained an access key Cursor had issued to partner Fireworks AI, used it to copy data containing leftover credentials, and on 29 October used one of those credentials to push a modified copy of the Claude Code extension into Cursor's extension marketplace. Up to 5,000 people may have received the tampered build.
Key points
- The attacker obtained an access key Cursor had issued to Fireworks AI (exposed by Fireworks, according to Cursor) and copied coding agent traces, mostly from 2025, from a shared storage bucket.
- A few traces contained unredacted credentials. These gave access to select SaaS tools used by Cursor, and on 29 Oct the attacker placed tampered Claude Code builds 2.1.312 and 2.1.313 in the marketplace cache.
- The extension sent host details to the attacker, added an SSH key, tried to open a reverse SSH tunnel and installed persistence. Anthropic was not compromised and did not build these versions.
- The attacker also reached archived non-production repositories, some support data and credentials users had sent to support. Cursor saw the last attacker activity on 1 Nov.
What happened
On 13 November 2026 Cursor explained how the incident began that led it to pull two tampered builds of the Anthropic.claude-code extension from its marketplace on 29 October. The starting point was an access key Cursor had issued to Fireworks AI, an AI inference provider. According to Cursor, Fireworks exposed the key during its October incident, which we covered in Fireworks AI breach hits Notion and Uber.
Technical details
The shared bucket held coding agent traces (mostly from 2025) from sessions without Privacy Mode. A few contained unredacted credentials. With them the attacker reached archived non-production repositories, some support tickets and emails, usage metrics, summaries of some agent conversations and credentials users had sent to support. Cursor says production systems, the app, official updates and signing keys were not affected.
Builds 2.1.312 and 2.1.313 (macOS and Windows) were available for about 11 hours on 29 Oct (2:27 am to 1:35 pm Pacific). According to Cursor, they collected host details including environment variable names and the Git email, added an attacker SSH key, tried to enable the SSH server and open a reverse tunnel, and installed persistence.
Who is affected
Up to 5,000 users who requested the affected builds from Cursor's marketplace on 29 Oct. Fewer than 0.01% of customers had API credentials exposed, and fewer than 0.05% of non-Privacy Mode users had agent traces exposed. Cursor notified those affected by email and in the app.
How to tell if you were compromised
Indicators from Cursor's notice (defanged, replace [.] with a dot):
Attacker IP: 134.209.82[.]225 (SSH tunnel tunnel@, TCP 22; POST to port 8099 /beacon)
SSH key: comment svc-telemetry
SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk
Tunnel key: SHA256:KOAU57+ZG6u3MG4WwzPzPX2O5YEe9qdOC1Bl+6MVyDQ
extension.js SHA-256 ecb58cb29078d9646ad1dcdbe51526f2d9cfea06be5fb2ab183e8ecc3ab6bf15
VSIX 2.1.312 darwin-arm64 fcbb327e38e61c4fec795e0d0591f52cd1014b813d20a5912104953c3a17c364
VSIX 2.1.312 win32-x64 3de1722901165769893bef57c72594e9a54459faa23f3091b136b79966edd081
VSIX 2.1.313 darwin-arm64 80402fdcad8799285fd93044976973ca000da7995fb93363f819bb8e4f84ee6c
VSIX 2.1.313 win32-x64 6786ec90cbd45949695bd7805c9fbb26cfc91bc5fc897de0183d9bcb9cb21731
macOS: ~/.config/.smt, ~/.config/.smtk, $TMPDIR/.stk_*, ~/Library/LaunchAgents/com.apple.sysmond.plist
Windows: %USERPROFILE%\.config\smt.cmd, scheduled task SysMon, ...\Startup\sysmon.cmd
Processes: ssh -N -R '*:0:localhost:22', autossh -M 0
Quick check on macOS: grep -qs svc-telemetry ~/.ssh/authorized_keys && echo FOUND. Requests to ifconfig.me and api.ipify.org mean nothing on their own. Turn the list into detections with our IOC extractor.
What to do
- Run Cursor's remediation script (macOS as the affected user, Windows from an administrator PowerShell). Cursor recommends it even if the check comes back clean.
- If you find the key, isolate the machine, remove the key, persistence and tunnel processes, and consider rebuilding it.
- From a separate, trusted machine, rotate credentials and sessions reachable from the affected host and every key Cursor listed for you.
- Don't paste secrets into coding agents or support tickets, and turn on Privacy Mode if its trade-offs work for you.
Timeline
- October 2026 - Fireworks AI handles two incidents; the key issued by Cursor is exposed.
- 29 Oct 2026 - tampered builds 2.1.312 and 2.1.313 live in Cursor's marketplace for about 11 hours, then pulled.
- 30 Oct 2026 - Cursor publishes indicators and remediation scripts and notifies users.
- 1 Nov 2026 - last observed attacker activity.
- 13 Nov 2026 - Cursor publishes the root cause and scope.
We verified these details on 16 Nov 2026 against Cursor's notices of 29-30 Oct and 13 Nov 2026.
FAQ
I use Claude Code directly from Anthropic or in VS Code. Am I affected?
According to Cursor, no. The tampered builds were served only from the cache behind Cursor's extension marketplace, not from Anthropic's official distribution or the Open VSX source. Cursor says no other extensions or versions were affected.
I didn't get a notice from Cursor. Do I need to do anything?
Cursor says that if you weren't notified, your data was not in the exposed traces. If you installed or updated the Claude Code extension in Cursor on 29 Oct, run the indicator checks anyway.
Why isn't uninstalling the extension enough?
According to Cursor, the malicious code set up standalone scripts that run at startup (a LaunchAgent on macOS, a scheduled task and a Startup folder script on Windows) and added an SSH key. Those remain after the extension is removed.
Sources
Spotted a mistake? Let us know and we will correct it openly.