Skip to content
Rush Commerce
Software & Dev4 min read

Cursor incident report: agent traces leaked credentials

The Cursor incident report ties its breach to a key exposed by Fireworks AI. Stored agent traces held unredacted credentials. Keep secrets out of agent chats.

The Cursor incident report is out, and it explains how the trojanized Claude Code extension got into the marketplace. The attack did not start at Cursor. It started with one access key that Cursor gave to Fireworks AI, its training and inference vendor. That key opened a bucket of old coding-agent traces. Some of those traces contained credentials in plain text. The lesson for a small team is simple: what you paste into an agent chat can be stored, shared with a vendor and stolen later.

What the Cursor incident report says

Cursor published its incident update on November 13. BleepingComputer reported it on November 14. The chain has four links:

  • The key. Cursor issued an access key to Fireworks. Cursor says Fireworks exposed it.
  • The bucket. The attackers used the key to copy data from a read-only storage bucket that Cursor shared with the vendor. BleepingComputer reports that the bucket held coding-agent traces, mostly from 2025, from users who did not have Privacy Mode on.
  • The credentials. Some traces contained unredacted API credentials. Those credentials opened more Cursor systems, which included several SaaS tools. The data there included archived non-production repositories, some support emails, usage metrics, and credentials that customers sent to support.
  • The marketplace. One stolen credential let the attackers publish extensions. On October 29 they put tampered copies of the Claude Code extension, versions 2.1.312 and 2.1.313, in the cache that serves Cursor's marketplace. The copies were live for approximately 11 hours. Cursor says at most 5,000 users received them.

The numbers are small in percent. Fewer than 0.05% of users had traces exposed. Fewer than 0.01% of customers had API credentials exposed. Cursor says the last attacker activity it saw was on November 1, and it was blocked.

Cursor says these were not affected: code and prompts from Privacy Mode users, production infrastructure, the Cursor app and its signed releases, and customer repositories. Cursor moved all inference away from the vendor, rebuilt the marketplace and redeployed production infrastructure.

We wrote about the trojanized extension on October 31. If a machine on your team got version 2.1.312 or 2.1.313, the cleanup steps in Cursor's notice still apply. Cursor says that uninstalling the extension alone may not be enough.

Why it matters for your business

The weak point was not the model. It was a log. An agent trace records everything: the prompt, the files the agent read, the command output. If a developer pastes a database password into the chat, or the agent prints an .env file, the secret is now in the trace. A year later that trace was in a vendor's bucket, and one key opened it.

Note the last item on the SaaS list too: credentials that customers sent to support. A support ticket is also a log.

What to do this week:

  1. Turn on Privacy Mode. Do it for every seat. Cursor says Privacy Mode code and prompts were not in the exposed data. Team admins can enforce it.
  2. Keep secrets out of agent context. Do not paste keys into chat. Add .env and credential files to .cursorignore. Use a secret manager and short-lived tokens.
  3. Rotate what the agent has seen. If a key was in a prompt, a terminal output or a support ticket, replace it.
  4. Scope keys that you give to vendors. One key, one bucket, read-only, with an expiry date. Cursor's key was read-only, and it was still sufficient.
  5. Ask each AI vendor two questions. Where are our traces stored? Which subprocessors can read them?

Key takeaways

  • Cursor's incident report traces the breach to an access key that Cursor says Fireworks AI exposed
  • The key opened stored agent traces; some traces held unredacted API credentials
  • Fewer than 0.05% of users had traces exposed; fewer than 0.01% of customers had credentials exposed
  • A stolen credential let attackers publish the tampered Claude Code extension on October 29, for approximately 11 hours
  • Privacy Mode data was not affected: turn it on, and keep secrets out of agent chats

Your agent's chat history is a credential store that nobody audits. We set up coding agents with enforced privacy settings, ignore rules for secret files and short-lived keys, so a stolen trace is worth nothing. See how we build agent workflows that keep secrets out of logs, or send us your setup and we will find the leaks with you.

Sources: Cursor incident update, BleepingComputer, Cursor extension security notice.

  • #cursor
  • #fireworks-ai
  • #security-incident
  • #ai-coding-agents
  • #credentials
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.