From a Partner’s Leaked Key to a Poisoned Coding Extension: A Timeline of How the Cursor Breach Unfolded
Cursor has published its incident report on the trojanized Claude Code extension that reached up to 5,000 developers in late October. The trail starts with an access key exposed at AI provider Fireworks AI. Here is the sequence, what was taken, and what was not touched.
Roughly two weeks after a booby-trapped copy of Anthropic’s Claude Code extension turned up in its marketplace, AI code editor maker Cursor has explained how the attacker got in. The answer is a supplier: an access key Cursor had handed to AI infrastructure company Fireworks AI was exposed, and the intruder used it as the first step in a chain that ended with malware on developers’ machines. The sequence below is drawn from Cursor’s Friday incident report as summarized by BleepingComputer, with additional detail from The Hacker News.
Before the attack: data shared with a vendor
Cursor kept a read-only storage bucket for exchanging data with Fireworks AI, which it used for model training and inference. The bucket held historical coding-agent traces, most of them from 2025, taken from sessions where users had not switched on Privacy Mode, the Cursor setting that keeps a user’s data out of training. Some of those traces still contained API credentials that had never been scrubbed out.
October: Fireworks AI is breached
Fireworks AI disclosed more than one security incident in October, including exposed API keys and tokens and a separate intrusion into development machines through a compromised third-party cloud provider. According to Cursor, the key it had issued to Fireworks was exposed in that fallout. The attacker used it to copy data out of the shared bucket.
Next: from traces to internal tools
The credentials buried in the traces gave the intruder a way into several of the software-as-a-service tools Cursor uses internally. From there, Cursor says, the attacker took:
- archived, non-production source code repositories, some of which contained business contact details;
- a limited amount of customer support information and email correspondence;
- credentials that customers had previously shared with Cursor’s support team.
Cursor puts the scale at fewer than 0.05% of users with coding-agent traces exposed and fewer than 0.01% of customers with API credentials compromised.
Detection and first response
Cursor’s monitoring flagged the unauthorized activity. The company switched off the Fireworks key, rotated exposed credentials and went looking for deeper access, and it temporarily shut down some internal and product services as a precaution, which disrupted some Cursor features. By then, though, the attacker already held a credential that could push content into Cursor’s extension marketplace.
October 29: the poisoned extension
Using that credential, the attacker placed tampered copies of the Claude Code extension, versions 2.1.312 and 2.1.313, straight onto the cache that serves Cursor’s VS Code extension marketplace. Anthropic’s genuine extension and the Open VSX registry Cursor relies on were not compromised. The modified code ran JavaScript that gathered system information, added an attacker-controlled SSH key, tried to switch on SSH access and set up persistence. Researchers found coding mistakes that likely stopped the backdoor from working on many Windows and Mac computers where SSH was not already enabled.
About 11 hours later: takedown
The malicious builds were live for roughly 11 hours. Cursor says at most 5,000 users who requested those specific cached versions on October 29 may have received them. It removed the builds, took the marketplace offline, worked with the attacker’s cloud hosting provider to take down their server, and notified affected users.
November 1 onward: quiet
Cursor says it has not seen further malicious activity since November 1. It has moved all inference traffic off Fireworks, cut Fireworks off from every data set it had shared, rebuilt the extension marketplace and redeployed its production infrastructure.
What Cursor says was not affected
Users with Privacy Mode enabled, customer code repositories, Cursor’s production environment, the Cursor app and its official updates, and other extensions and plugins from its marketplace, according to the company.
The bigger pattern
This is one of a string of attacks on AI developer platforms that security firm Wiz has labeled “Black October,” in which stolen keys for AI services are used to run up inference charges or worse. Railway, Northflank, Lightning AI, OpenCode and Warp have all been named as affected. The Fireworks AI incidents also reached Notion, and The Hacker News notes that Uber has confirmed driver’s license data was taken through an unnamed service provider that TechCrunch reported was Fireworks AI.
What this means for your company
Even a business that has never heard of Fireworks AI can learn from this chain, because every link in it exists in ordinary companies:
- Data shared with vendors is your exposure too. Keep an inventory of what you hand to suppliers, where it sits and which keys open it. Remove shared data sets that are no longer needed.
- Secrets do not belong in logs, chats or tickets. Scan for passwords and API keys in support threads, AI tool histories and shared files, and rotate anything that has been pasted somewhere it should not be.
- Check AI tool privacy settings. If staff use AI coding or writing assistants, decide as a company whether their sessions may be used for training, and set it centrally where the tool allows.
- Treat extensions as software you install. Limit which editor and browser extensions are allowed, and have someone watch for vendor notices about compromised versions.
- Watch for new SSH keys and remote access. If anyone on your team used Cursor with the Claude Code extension around October 29, follow Cursor’s remediation guidance and check those machines for unfamiliar SSH keys and services.
If your developers or power users rely on AI tools, now is a good time to map which vendors hold your data and keys. Reach out to Layer 7 Data Solutions to discuss how we’ve helped companies similar to yours.
