Cursor Security Incident: 7 Costly Lessons for AI Coders

A key exposed by Fireworks AI let an attacker copy old Cursor agent traces, harvest credentials and plant a malicious Claude Code build for 11 hours. The timeline, scope and seven lessons for teams.

This breakdown is based on Cursor’s incident update and security notice, Fireworks AI’s bulletin and reporting by security outlets. I have no inside knowledge of the investigation, both companies say their findings may change, and every detail is attributed.

The Cursor security incident started somewhere else entirely. On 23 October 2026, a third-party cloud provider hosting Fireworks AI’s research machines was compromised. An access key that Cursor had issued to Fireworks was exposed, and an attacker used it to copy old coding-agent traces, harvest credentials hidden inside them and, on 29 October, place a malicious build of Anthropic’s Claude Code extension in the cache serving Cursor’s extension marketplace for about 11 hours.

Cursor says the incident is contained, and the numbers affected are small. But the chain is the lesson. A vendor’s cloud provider leaked a key, the key opened a bucket of logs, the logs contained secrets, and the secrets opened a software supply chain. If you use AI coding tools at work, every link in that chain exists in your setup too.

Table of Contents

Cursor security code editor on a laptop
The editor was not the weak link. Secrets sitting in old agent traces were, and that is true of every AI coding tool.

Cursor Security Incident Timeline

Here is the sequence as Cursor and Fireworks have described it, compiled by Unite.AI.

Date (2026) Event
23 October A third-party cloud provider hosting Fireworks AI’s R&D machines has an incident
24 October An unauthorised party reaches Fireworks development machines; Fireworks detects it the same day
25 October Last successful unauthorised access Fireworks has detected on its customer platform
29 October The attacker places unauthorised Claude Code extension builds on the cache serving Cursor’s marketplace, for about 11 hours
30 October Cursor publishes indicators of compromise and begins notifying affected users and admins
1 November Last attacker activity Cursor observed, which was blocked
9 November Fireworks expands its account to include account, billing and some dataset information
13 November Cursor publishes its full incident update

The root cause, per Cursor, was a Cursor-issued access key that Fireworks exposed. It allowed data to be copied from a read-only bucket the two companies used to share data. Fireworks had been one of Cursor’s inference providers.

What the Attacker Accessed

Cursor’s update lists what was reached, and it is a useful map of where secrets hide in an AI coding workflow.

  • Coding-agent traces. A limited set, mostly from 2025, from sessions not using Privacy Mode that Cursor could use for training. Some contained unredacted credentials.
  • Credentials found in those traces. Used to reach select SaaS tools and then archived, non-production source repositories, some with business contact information.
  • Other data. Limited support data and emails, product usage metrics and contact details, summaries of some agent conversations, and credentials that users had sent to support.

Cursor says production systems, live traffic, current customer data, Privacy Mode prompts and code, customer repositories, official app updates and release-signing keys were not affected. Open VSX never carried the malware, and Anthropic’s own product was not vulnerable. Cursor stresses that Anthropic did not publish the malicious builds.

On scale, Cursor says fewer than 0.01% of customers had API credentials exposed and fewer than 0.05% of non-Privacy Mode users had traces exposed. At most 5,000 users who requested specific cached versions may have received the malicious extension during its roughly 11-hour window.

Were You Affected by the Cursor Security Incident?

Cursor says it notified affected users and administrators directly, and sent admins lists of specific keys to rotate. If you did not get an email, you were most likely not in the affected group. Still, check three things.

The extension version. Cursor’s security notice names Anthropic.claude-code versions 2.1.312 and 2.1.313 for Windows and macOS, with an exposure window of roughly 2:27am to 1:35pm Pacific time on 29 October, according to AIToolRanked’s summary of the notice. If you installed or updated the Claude Code extension inside Cursor that day, check your installed version and the indicators of compromise Cursor published.

Your Privacy Mode setting. Exposed traces came from sessions without Privacy Mode. If you never turned it on, assume anything you pasted into an agent session in 2025 could have been in scope.

Secrets you shared with support. If you ever sent an API key or password to Cursor’s support team, rotate it now, whether or not you were notified.

Lesson 1: Your Vendors’ Vendors Are Your Attack Surface

Cursor was not breached first. Its inference provider’s cloud host was. That is two steps removed from Cursor and three from the developers whose secrets ended up exposed.

Every AI tool you use sends data to model providers, inference hosts, logging services and analytics platforms. Each one holds keys or data that can unlock the next. You cannot audit all of them, but you can limit what flows through them, which is what the rest of these lessons are about.

Cursor’s response is instructive: it moved all inference traffic off Fireworks, deactivated the vendor key and cut Fireworks’ access to every shared dataset. Ask your own AI vendors which subprocessors hold your data, and what happens to their keys if one of them is compromised.

Lesson 2: Cursor Security Traces Were Full of Secrets

The most important detail in the Cursor security incident is that coding-agent traces contained unredacted credentials. That should surprise nobody who has used an agent. You paste an error message with a connection string in it. The agent reads a .env file to debug a config problem. A tool call echoes an API key back into the log.

All of that ends up in a trace. Traces get stored, shared with partners for training or debugging, and kept for months. A trace from 2025 was still useful to an attacker in late 2026.

The practical fix is to keep secrets out of what agents see. Use a secrets manager and short-lived tokens rather than long-lived keys in files. Our guide to context engineering covers how to give agents the context they need without handing them your credentials.

Lesson 3: Extension Marketplaces Are a Cursor Security Weak Point

The malicious Claude Code build did not come through Anthropic. It came through the cache serving Cursor’s own marketplace, using a stolen credential. Developers who trusted the marketplace got a trojanised extension for about 11 hours.

Extensions for AI coding tools are especially dangerous because they often run with broad access: your files, your terminal and sometimes your cloud credentials. After the incident, Cursor validated its other marketplace extensions and then reviewed and rebuilt the store.

For your team, the takeaway is to treat extensions like dependencies. Pin versions, review updates for tools with high privileges and prefer extensions installed from the publisher’s official channel. The same thinking applies to MCP servers, which we cover in our guide to prompt injection and agent security.

Lesson 4: Cursor Security Shows Why You Rotate Everything

Cursor rotated every credential it identified as exposed, then rotated all credentials in use internally as a precaution, and asked third-party services to revoke exposed keys. That second step is the one most teams skip, and it is the one that catches the secret you did not know was exposed.

It came at a cost. Cursor says it temporarily disabled more internal services than strictly needed, which disrupted some features before they were restored. That is the right trade. A few hours of disruption is cheaper than a second breach through a key you missed.

If you have never rotated your team’s keys, use this incident as the excuse. Make a list of every API key your developers use with AI tools, give each one an owner and an expiry date, and rotate the lot on a schedule.

Lesson 5: Cursor Security Now Scans Data, Not Just Code

Most teams scan code repositories for leaked secrets. Far fewer scan logs, traces, support tickets or shared datasets. Cursor’s post-incident changes included adding secret scanning to repositories and data sources, expanding network restrictions and redeploying production infrastructure with external forensic help.

The phrase “and data sources” is the one to copy. Wherever agent conversations, logs or support threads are stored, run the same secret detection you already run on code. A credential in a log is just as usable as one in a commit.

Lesson 6: Privacy Mode Is a Cursor Security Setting

Every exposed trace in the Cursor security incident came from sessions that were not using Privacy Mode. Cursor says Privacy Mode prompts and code were not affected.

Many developers think of privacy settings as being about training: do I want my code used to improve the model? This incident shows they are also about blast radius. Data that is not retained cannot be stolen later. If your AI coding tool offers a zero-retention or privacy mode, turn it on for any work involving client code or credentials, and enforce it at the organisation level if you can.

Lesson 7: AI Coding Tools Cut Both Ways

This is not the first time Cursor has appeared in a security story in 2026. In August, Gambit Security reported that Russian-speaking attackers used Cursor’s AI agent during intrusions at seven companies, which Reuters could not fully verify. The Cloud Security Alliance described the agent’s role as an assistant to a human operator rather than an autonomous attacker.

Researchers also disclosed vulnerabilities in several AI coding agents this year, including sandbox escapes reported in July affecting Cursor, OpenAI’s Codex, Google’s Gemini CLI and Antigravity. Cursor is not uniquely weak. It is simply very popular, which makes it a target.

The broader point: AI coding tools are now critical infrastructure for many teams. They deserve the same security review as your CI system or your cloud console. Our roundup of AI coding assistants in 2026 covers the privacy and admin controls each one offers.

What Cursor Security Changes Mean for You

Cursor listed a series of permanent changes alongside the clean-up, and several of them change what you can expect from the product. It expanded network restrictions, added secret scanning to repositories and data sources, redeployed its production infrastructure and brought in external forensics experts. It also validated every other extension in its marketplace before rebuilding the store.

For users, the most visible effect is likely to be stricter handling of anything that looks like a credential, and possibly more friction when installing extensions. That is a good trade. Cursor security is now being judged by enterprise buyers who will read this incident report closely, and tighter defaults are the fastest way to win back their confidence.

It is worth asking whether Cursor’s new owner changes the picture. Cursor has been owned by SpaceX since its acquisition of parent company Anysphere, and OpenAI ended direct model access on 12 November citing its own concerns about that ownership. Neither the incident report nor the security coverage links the breach to the acquisition. But for procurement teams, vendor ownership and vendor security now sit in the same risk review.

The healthiest response is not to drop the tool. It is to stop treating any AI coding tool as a black box. Ask each vendor for its subprocessor list, its retention defaults and its incident history, and write the answers down. Cursor security got better this month because it had to. Your review process can make sure the others follow.

A Checklist After the Cursor Security Incident

If you are Start with Why Link
A Cursor user who installed Claude Code on 29 October Check your extension version against the IOCs Versions 2.1.312 and 2.1.313 were named Cursor’s incident update
A team admin Rotate the keys on Cursor’s list, then all AI-tool keys Exposed credentials came from traces and support data BleepingComputer report
Any AI coding tool user Turn on privacy or zero-retention mode Every exposed trace came from non-Privacy Mode sessions AI coding assistants
A security lead Add secret scanning to logs and data stores Credentials in traces are as usable as those in code Agent security guide
A Fireworks customer Read Fireworks’ own bulletin Its investigation is ongoing and findings may change Unite.AI summary

Related on this site: for more context, see Copilot AI Credits Explained and GPT-6.1 Sol.

FAQ About the Cursor Security Incident

What happened in the Cursor security incident?
An access key Cursor had issued to Fireworks AI was exposed after a breach at Fireworks’ cloud provider. An attacker used it to copy old agent traces, extract credentials, reach some internal tools and place a malicious Claude Code extension build in Cursor’s marketplace cache for about 11 hours on 29 October 2026.

Was Anthropic’s Claude Code hacked?
No. Cursor says Anthropic’s product was not vulnerable and Anthropic did not publish the malicious builds. The tampered copies were placed in Cursor’s own marketplace cache.

How many users were affected?
Cursor says at most 5,000 users may have received the malicious extension, fewer than 0.01% of customers had API credentials exposed, and fewer than 0.05% of non-Privacy Mode users had traces exposed.

Is my code safe?
Cursor says customer code repositories, current customer data and Privacy Mode prompts and code were not affected. Archived, non-production repositories belonging to Cursor itself were accessed.

Is it still safe to use Cursor?
Cursor says the incident is contained, with the last attacker activity blocked on 1 November. Turn on Privacy Mode, keep secrets out of agent sessions and follow any rotation instructions you received.

Did Cursor stop using Fireworks?
Yes. Cursor moved all inference traffic off Fireworks, deactivated the vendor key and removed Fireworks’ access to shared datasets.

Verdict

The Cursor security incident is contained and affected a small fraction of users, but its shape is what every team using AI coding tools should study. A breach two vendors away turned into stolen secrets, then into a poisoned extension, because agent traces held credentials and those traces were retained and shared.

The caveat is that both investigations are still open, and Fireworks has already expanded its account once. The numbers above may change, and the full root cause at the cloud provider has not been published.

Your next step: this week, turn on privacy or zero-retention mode in every AI coding tool your team uses, rotate any key that has ever appeared in an agent session or support ticket, and pin the versions of high-privilege extensions.

Sources and Further Reading

On this site, read prompt injection and agent security for the wider threat model, AI coding assistants in 2026 for tool-by-tool privacy controls, and local AI agents if you want to keep code entirely on your own machine.