Cursor's breach started with saved AI coding sessions that held live keys. Estimators building bid tools should keep logins out of the session
Cursor says the attack that planted a backdoored extension in its marketplace began with stored records of AI coding sessions, a few of which contained working credentials. For an estimator or sub building internal tools, the fix is a privacy setting, a habit, and a short list of keys to rotate.
If someone at your company builds bid or back-office tools with an AI coding assistant, anything the assistant reads during a session can end up in the vendor's stored records, including passwords and API keys. Cursor confirmed on November 13 that its recent breach started exactly there: an attacker copied saved records of AI coding sessions, found working credentials inside a few of them, and used those to plant a backdoored extension in Cursor's marketplace.
We covered the backdoored extension when it surfaced. At the time nobody knew how it got there. Now there's an answer, and it changes what a contractor should do.
What did Cursor say happened?
Per Cursor's investigation update:
- Cursor shared a read-only storage bucket with Fireworks AI, a vendor that ran AI models for it. Cursor says Fireworks exposed the access key, and an attacker used it to copy the bucket.
- The bucket held "coding agent traces," which are records of AI coding sessions. They were mostly from 2025 and came from sessions not in Privacy Mode. Cursor says it was permitted to use them for model training.
- Some traces contained unredacted credentials. Cursor says less than 0.01% of customers had API credentials exposed.
- The attacker used those credentials to reach some of Cursor's internal software tools, then on October 29 placed malicious copies of the Claude Code extension, versions 2.1.312 and 2.1.313, on the cache that serves Cursor's marketplace. They were up for about 11 hours. At most 5,000 users may have received them.
Cursor says Privacy Mode prompts and code, customer code repositories, its production systems and the Cursor app itself were not affected. It has moved all inference off Fireworks. One caution: the attribution to Fireworks is Cursor's account. iFynx notes that Fireworks' own bulletins do not mention Cursor. We wrote about Fireworks' disclosure last week.
Why does this matter to an estimator or a sub's back office?
A coding agent doesn't only see what you type. It reads files in the project folder, runs commands, and sees the output. That is how a credential lands in a session record without anyone pasting it on purpose.
Picture an estimator building a script that pulls committed costs from the accounting system into a bid-day spreadsheet. The API key sits in a config file in the same folder. The agent opens the file to debug a connection error. The key is now part of the session. If that session is stored, the key is stored with it.
Cursor's data-use page spells out the difference. With Privacy Mode off, Cursor "may use and store codebase data, prompts, editor actions, code snippets" to improve its features and train models. With it on, Cursor says customer data is not used for training.
What should you change this week?
| Do this | Why |
|---|---|
| Turn on Privacy Mode for everyone who codes for you, including outside consultants | Cursor says Privacy Mode sessions were not in the exposed data |
| Keep live keys out of the folder the agent works in; use a test account or a read-only key while building | The agent can read any file it can open |
| Rotate any key that was ever pasted into a session or sat in an agent-readable file during 2025 | Deleting the file does not cancel the key; rotate it in the system that issued it |
| Rotate anything you ever emailed to Cursor support | Cursor says credentials users sent to support were among the data accessed |
| Check for extension versions 2.1.312 and 2.1.313 installed October 29 | Cursor's notice says uninstalling alone may not be enough |
The same habits apply to any AI coding tool, not only Cursor. Ask each vendor whether sessions are stored, for how long, and which outside companies can read them.
What are the limits here?
The numbers are small and they are Cursor's own. Cursor says fewer than 0.05% of users outside Privacy Mode had traces exposed, that all affected customers were contacted, and that anyone not notified was not in the data. Nothing in the disclosure points to a construction company. If nobody on your payroll or your consultant list writes code with an AI tool, this does not reach you.
It does show that the session record is a copy of whatever the assistant touched. A bid tool that takes an afternoon to build can leave an accounting key in a vendor's stored records long after the tool is finished.
Takeaway: Have whoever builds your internal tools turn on Privacy Mode today, list every key their project folders have held, and rotate those keys at the source. Then make "test credentials only while building" a written rule before the next tool gets started.
Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com and forward this to whoever builds your spreadsheets and scripts.
- What caused the Cursor security incident?
- In a November 13, 2026 update, Cursor said an attacker obtained an access key Cursor had issued to its inference vendor Fireworks AI and used it to copy stored coding agent traces from a shared storage bucket. A small number of those traces contained unredacted credentials, which the attacker used to reach Cursor's internal tools and plant malicious copies of the Claude Code extension in its marketplace cache on October 29.
- Does Cursor save what I type into its AI coding agent?
- It can. Cursor's data-use page says that with Privacy Mode off it may use and store codebase data, prompts, editor actions and code snippets to improve its AI features and train its models. With Privacy Mode on, Cursor says customer data is not used for training. Cursor says prompts and code from Privacy Mode sessions were not affected by this incident.
- Was my company's data exposed in the Cursor breach?
- Cursor says it has contacted all affected customers and that anyone who was not notified was not in the exposed traces. It puts exposed API credentials at less than 0.01% of customers and exposed traces at fewer than 0.05% of users not in Privacy Mode. Customer code repositories and production systems were not affected, according to Cursor.
- What should a contractor do if staff use Cursor to build internal tools?
- Turn on Privacy Mode, keep accounting, project-management and cloud-storage keys out of files and prompts the coding agent can read, and rotate any credential that was ever pasted into a session or emailed to Cursor support. Anyone who installed Claude Code extension 2.1.312 or 2.1.313 from the Cursor marketplace on October 29 should follow Cursor's full remediation notice.