Cursor, the popular AI code editor, has explained how a breach at one of its AI suppliers ended with a backdoored copy of Anthropic’s Claude Code extension being handed to its users. In an incident update published on Friday (US time), Cursor said an access key it had issued to AI inference company Fireworks was exposed, and that an attacker used it to reach old coding-agent records containing unredacted credentials. Those credentials led further into Cursor’s systems, and eventually to its extension marketplace.
The malicious extension was live for about 11 hours on 29 October, and Cursor says at most 5,000 users may have received it. If you or your developers use Cursor, here’s what happened, who was affected, and what to check.
How the breach unfolded
According to Cursor’s update, the chain of events ran like this:
- An attacker got hold of an access key Cursor had issued to Fireworks, which Cursor said was “exposed by Fireworks”. The key gave read-only access to a storage bucket Cursor used to share data with the vendor.
- That bucket held coding-agent traces, mostly from 2025, from sessions where users had not turned on Cursor’s Privacy Mode and which Cursor was permitted to use for model training. Some traces contained API keys and other credentials that had not been redacted.
- Using those credentials, the attacker reached some of the software-as-a-service tools Cursor uses. From there they accessed archived, non-production Cursor source code repositories, limited customer support data and emails, product usage metrics and contact details, summaries of some coding-agent conversations, and some credentials customers had sent to Cursor’s support team.
- On 29 October, using another stolen credential, the attacker placed tampered copies of Anthropic’s Claude Code extension, versions 2.1.312 and 2.1.313, directly into the cache that serves Cursor’s VS Code extension marketplace.
Cursor stresses that “Anthropic did not publish this copy, and this was not a vulnerability in Anthropic’s product.” It says the upstream marketplace, Open VSX, never carried the malware. Microsoft’s threat researchers, who analysed the malicious package, separately confirmed it was never distributed through Microsoft’s Visual Studio Marketplace or Anthropic’s npm package.
Fireworks has disclosed its own incidents. In a later security update, it said it had found impacts to account and billing information and, for a limited set of accounts, to secrets, datasets and incomplete models, and linked the activity to a financially motivated group selling stolen access to large language models, known as “LLMJacking”. BleepingComputer reports that other AI developer platforms were hit in a wave of attacks that security firm Wiz has dubbed “Black October”.
What the malicious extension did
Cursor’s original security notice says the extension was available between about 2:27am and 1:35pm Pacific time on 29 October, which is roughly 8:27pm on 29 October to 7:35am on 30 October in Sydney (our conversion). Only the Windows and macOS builds were affected. When the extension loaded, the injected code:
- collected host details such as the computer name, username, operating system, home folder path, the names of environment variables, Git email address and public IP address, and tried to send them to the attacker
- added an attacker-controlled SSH key, labelled
svc-telemetry, to the user’s authorised keys file, and tried to turn on the computer’s SSH service and open a reverse SSH tunnel to the attacker’s server - tried to make that tunnel survive restarts, using a LaunchAgent on macOS and a scheduled task or Startup-folder script on Windows
There is some good news. Cursor says that, because of errors in the malicious code and because enabling SSH needs administrator rights the extension usually didn’t have, the tunnel step would have failed on most Macs and Windows PCs, so the attacker could not log in that way. Cursor also says it worked with the hosting provider to take the attacker’s server offline. But Microsoft warns that “the loader may still write keys and persistence artifacts even when tunnel access fails”, so an affected machine can still carry leftover backdoor pieces.
Who was affected, according to Cursor
- Malicious extension: at most 5,000 users who requested the affected cached versions on 29 October may have received the tampered build. That is a potential distribution figure, not a count of compromised machines.
- Coding-agent traces: fewer than 0.05% of users had traces exposed, all from sessions not using Privacy Mode. Cursor says it has notified each of them, and if you didn’t get a notice your data was not in those traces.
- Credentials: fewer than 0.01% of customers had API credentials exposed. Cursor says it emailed team admins lists of specific keys to rotate.
- Not affected, Cursor says: prompts and code from users with Privacy Mode on, customer code repositories, Cursor’s production systems, and the Cursor app and its official updates, which ship from a separate system the attacker never accessed. Cursor says no release signing keys were exposed.
Cursor says the incident is contained and that the last attacker activity it saw, on 1 November, was blocked. It has moved all inference traffic away from Fireworks, rotated credentials across its internal systems, rebuilt its extension marketplace, added secret scanning and redeployed its production infrastructure. The company apologised and wrote that “we can and must do better”. Cursor hasn’t said how many affected users are in Australia, and these figures are its own, not independently verified.
What Australian developers and businesses should do now
For Australian developers and software teams using Cursor, this incident fits a pattern Australia’s cyber agency has been warning about. The Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC) has an active alert on the targeting of online code repositories and software packages, and its Annual Cyber Threat Report 2025–26 says it has tracked this activity since September 2025. “The compromise of trusted software packages presents a significant and ongoing risk for organisations,” the alert says.
- Check for the key. If anyone on your team installed or updated the Claude Code extension in Cursor on a Mac or Windows PC around 29–30 October (Australian time), look for an entry labelled
svc-telemetryin their SSH authorised keys file. Cursor’s security notice has one-line checks for macOS and Windows, remediation scripts and a full list of indicators. - Don’t just uninstall. In Cursor’s words: “Uninstalling the extension alone may not be enough.” The payload drops separate persistence scripts, so remove the key, the startup items and any tunnel processes, and consider rebuilding the machine.
- Rotate from a clean device. Rotate passwords, API keys and session tokens that were reachable from an affected machine, using a trusted computer, as Cursor advises. Do the same for any keys Cursor told your admins to rotate.
- Stop pasting secrets into AI tools. The attacker’s route deeper into Cursor ran through credentials sitting in old AI coding-agent logs. ASD’s ACSC guidance on managing cryptographic keys and secrets is a good place to start on keeping keys out of code, chat and logs, and using short-lived, least-privilege credentials.
- Review Privacy Mode. If your organisation uses Cursor, check whether Privacy Mode is turned on for work that touches sensitive code or credentials.
- Report it. If you find evidence of compromise, Australian organisations and individuals can report it to ASD’s ACSC.
Supplier risk has been a theme this week. Expedia said its attackers got in through a third-party vendor, and attackers hijacked ScreenConnect through a forgotten add-on used by IT providers.
Disclosure: The Tech Boom uses Grok Bot, an AI assistant made by Cursor, in its editorial workflow. Cursor lists Grok Bot among its products and says it was not affected by this incident. Cursor had no input into this article.
Our read
Our read: The most uncomfortable detail here isn’t the malicious extension, which Cursor says was removed after about 11 hours and mostly failed to open its tunnel. It’s that ordinary developer sessions, kept for model training, quietly became a store of working passwords and API keys, and that store sat with a third-party vendor. AI coding agents see everything a developer pastes, runs and prints, so their logs need to be treated with the same care as a password vault. Cursor has been fairly detailed in its disclosure and its cleanup steps, which helps. But teams adopting AI coding tools should now ask every vendor where agent traces go, who else can see them, and how quickly secrets are scrubbed, before something similar happens again.
Featured image: a developer writing code on a laptop (generic illustration; not Cursor, not the affected extension, and not connected to this incident). Photo: Tirza van Dijk, via Wikimedia Commons (originally Unsplash), released under CC0 1.0; resized and converted to PNG.




