At the start of the month we wrote about tampered Claude Code extensions on Cursor's marketplace. On 13 November, Cursor published an update on that incident after about two weeks of investigation. The extension turns out to be only part of the story. The attack started at one of Cursor's vendors, and what the attacker copied first was coding agent traces.
For most teams, the takeaway is not about Cursor. It is about a kind of data nobody thinks of as sensitive until it leaks: the record of what your coding agent read, ran and wrote.
What Cursor says happened
According to Cursor, an attacker obtained an access key that Cursor had issued to Fireworks, its inference vendor. Cursor says the key was exposed by Fireworks. With it, the attacker copied data from a read-only storage bucket Cursor used to share data with Fireworks.
That bucket held coding agent traces, mostly from 2025, from sessions that were not in Privacy Mode and that Cursor was allowed to use for model training. Some traces contained unredacted credentials. Cursor says traces were exposed for fewer than 0.05% of users not in Privacy Mode, and that less than 0.01% of customers had API credentials exposed.
Using credentials found that way, the attacker reached some of Cursor's SaaS tools. Cursor lists archived, non-production source code repositories, limited support data and emails, usage metrics and contact details, summaries of some agent conversations, and some credentials users had sent to Cursor's support team. On 29 October, the attacker also used a stolen credential to place malicious copies of the Claude Code extension (versions 2.1.312 and 2.1.313) on the cache serving Cursor's extension marketplace. Cursor says they were up for about 11 hours and could have reached at most 5,000 users. It stresses that Anthropic did not publish those copies and that the upstream Open VSX marketplace never had the malware.
Cursor says Privacy Mode prompts and code, customer code repositories, production systems and the Cursor app itself were not affected, and that no release signing keys were exposed. It has moved all inference off Fireworks, rotated credentials, rebuilt the marketplace and added secret scanning. It says the last attacker activity it saw was on 1 November and that it believes the incident is contained.
Fireworks has its own account. As Unite.AI summarizes it, a third-party cloud provider hosting Fireworks' R&D machines had an incident on 23 October, an unauthorized party reached Fireworks development machines on 24 October, and Fireworks says its investigation is still ongoing.
Why this matters beyond Cursor
Three things stand out, and none is specific to one tool.
Agent traces carry secrets. An agent reads config files, prints environment variables while debugging, and gets credentials pasted into chat. All of that can end up in a trace. Cursor's numbers are small in percentage terms, but the point stands: a log of agent sessions can be a credential store nobody meant to create.
Your data travels further than your vendor. The breach didn't start at the company the customers chose. It started at a company that company chose, and arguably at that company's cloud provider before that. Every coding tool sends work to model providers and other subprocessors. Your exposure is the weakest link in that chain.
Settings decide what is kept. Users in Privacy Mode were not in the exposed traces, according to Cursor. Whether a session is retained, and whether it can be used for training, is often a setting a developer picked once, or never looked at.
What to check on your team
- If you use Cursor, read the notice. Cursor says it contacted affected customers directly and that team admins received a list of keys to rotate. If anyone installed or updated the Claude Code extension to 2.1.312 or 2.1.313 from Cursor's marketplace in the window, follow the steps in Cursor's 29 October notice. Cursor warns that uninstalling the extension alone may not be enough.
- Check data settings for every coding tool. For each agent your team uses, find out whether sessions are retained, for how long, and whether they can be used for training. Where the tool offers an organization-wide privacy or zero-retention setting, decide on it once for the team instead of leaving it to each person.
- Ask about subprocessors. Your vendor's security page or data processing terms should list who else handles your data. If you can't find it, ask. It is a fair question after this month.
- Keep secrets out of agent sessions. Don't paste keys into chat. Keep production credentials out of the agent's working directory and use short-lived, narrowly scoped tokens for anything an agent needs. A secret that never enters a session can't leak from a trace.
- Treat support tickets the same way. Some exposed credentials had been sent to Cursor's own support team. When you share logs with any vendor, scrub them first.
- Have a rotation plan. If a vendor emails tomorrow with a list of exposed keys, you want to know who rotates them and how long it takes. Practise on one low-risk key.
The takeaway
Cursor's report is detailed, and it gives customers specific actions, which is what a good incident update should do. The lesson for everyone else is that coding agents produce a new kind of record, full of code, commands and sometimes secrets, and that record lives with your vendors and their vendors. Know where it goes, keep what goes into it clean, and choose the retention settings on purpose.
Sources: Cursor, “An update on our recent security incident” (13 November 2026), Unite.AI, “Cursor Traces Breach to Access Key Exposed by Fireworks AI” (13 November 2026).