AI & Cybersecurity

One Leaked AI Key Put a Fake Claude Code Extension in Front of Up to 5,000 Cursor Users

Aditya Kumar JhaAditya Kumar JhaLinkedInAmazon·November 15, 2026·9 min read

Cursor says a key exposed by Fireworks AI led to a malicious Claude Code extension for up to 5,000 users. Who was hit and what to rotate.

On Friday, November 13, 2026, Cursor published an incident report saying attackers used an access key that Cursor had issued to a vendor, Fireworks AI, to break into data and then slip trojanized copies of Anthropic's Claude Code extension into Cursor's extension marketplace. Cursor estimates up to 5,000 users could have received the bad build during an 11-hour window on October 29. The story is a clean case study in how AI tooling is attacked: not through the model, but through keys, vendors and add-ons. (Sources: BleepingComputer, The Hacker News)

Insight

Quick summary: Per Cursor's account, an attacker obtained a key that Fireworks AI exposed, copied agent session traces from a shared read-only storage bucket, found some unredacted credentials in them, and used those to reach more systems. On October 29 a trojanized Claude Code extension (versions 2.1.312 and 2.1.313) sat in Cursor's marketplace cache for about 11 hours. Cursor says Privacy Mode users, customer code repositories and its production environment were not affected, and it saw no further malicious activity after November 1. If you installed those versions, or ever pasted credentials into Cursor support, rotate them.

How the attack worked, according to Cursor

Cursor's report, as relayed by BleepingComputer and The Hacker News, describes a chain with four links. The attacker got an access key Cursor had issued to Fireworks AI, which Cursor says Fireworks exposed. The attacker used it to copy data from a read-only storage bucket that Cursor used to share information with Fireworks. Some of the records in that bucket contained unredacted API credentials. Those credentials opened doors to additional systems. Note that this is Cursor's version of events, and we found no direct statement from Fireworks about the Cursor key in the coverage we reviewed or in Fireworks' own security bulletins. (Sources: Cursor security incident update, Fireworks security bulletins, BleepingComputer, The Hacker News)

What was in the bucket matters. The Hacker News says it held limited coding agent traces from sessions that were not in Privacy Mode, mostly from 2025 and permitted for model training. Traces are records of what an AI coding agent saw and did, and developers sometimes paste secrets into prompts or the agent reads them from environment files. That is how a storage bucket full of text turned into a credential source.

The numbers Cursor and reporters gave

  • Up to 5,000 users could have received the malicious extension, because they requested the affected cached versions during the window.
  • About 11 hours: how long the malicious versions were available on October 29 before removal.
  • Fewer than 0.05% of users had traces exposed, per BleepingComputer.
  • Fewer than 0.01% of customers had API credentials compromised, per both outlets.
  • No additional malicious activity since November 1, 2026, per Cursor via The Hacker News.

The two outlets describe the percentages slightly differently, with BleepingComputer attaching the 0.05% figure to exposed traces and both attaching 0.01% to credentials. Those are Cursor's own figures, and no independent party has verified them. (Sources: BleepingComputer, The Hacker News)

What the fake extension did, and what it did not

According to BleepingComputer, attackers modified the legitimate Claude Code extension to run malicious JavaScript. The code collected system information, installed an attacker-controlled SSH key, tried to enable SSH access and set up persistence. Security researchers found coding errors that likely stopped the backdoor from working on many Windows and macOS systems where SSH was not already enabled. That is reassuring but not a guarantee, so anyone who installed the affected versions should treat the machine as potentially exposed. Both outlets say Anthropic's legitimate extension and Open VSX, the marketplace Cursor uses, were not compromised. The tampering happened in Cursor's cache. (Source: BleepingComputer)

ItemStatus per CursorSource
Privacy Mode users' code and promptsNot reported as affectedBleepingComputer
Customer code repositoriesNot reported as affectedBleepingComputer, The Hacker News
Production environment and official app updatesNot reported as affectedBleepingComputer, The Hacker News
Archived non-production source repositoriesAccessed, some with business contact informationThe Hacker News
Support emails and credentials sent to supportAccessed in limited formThe Hacker News, BleepingComputer
Claude Code extension 2.1.312 or 2.1.313 from the marketplaceMalicious. Users notified and helped with credential revocationThe Hacker News

Cursor's response

Cursor says it deactivated the Fireworks key, revoked and rotated exposed credentials, temporarily shut down some internal and product services, took the marketplace offline and rebuilt it, worked with the hosting provider to remove the attacker's server, and moved all inference traffic off Fireworks. It also cut Fireworks' access to shared datasets. The service disruption users noticed was part of that precaution. (Sources: The Hacker News, BleepingComputer)

The bigger pattern: a vendor in the middle

Fireworks AI is a hosting provider for AI models. BleepingComputer reported on November 10 that Fireworks disclosed a second breach after an October 23 compromise at an unnamed third-party cloud provider that hosted its R&D machines. Data accessed included resource metadata, billing and shipping addresses, invoice history, and partial model weights and datasets for some customers. Fireworks said payment card numbers were not accessible and that an earlier October incident was linked to a financially motivated group that sells unauthorized model access, known as LLMjacking. (Source: BleepingComputer)

The Hacker News adds that Notion confirmed its data was stolen in the Fireworks hack, and that Uber confirmed thousands of drivers' license records were stolen after a service provider was compromised. The Hacker News attributes the identification of that provider as Fireworks to TechCrunch, and Uber has not named it, so treat that link as reported and not confirmed by Uber. BleepingComputer says Wiz has labeled a wave of attacks on AI developer platforms "Black October," and lists Railway, where stolen secrets were used to generate about $740,000 in fraudulent AI charges, plus Northflank, Lightning AI, OpenCode and Warp. (Sources: The Hacker News, BleepingComputer)

What to do this weekend

  • Check your extensions. If you use Cursor, see whether Claude Code 2.1.312 or 2.1.313 was installed from the marketplace around October 29. Cursor says affected users were notified, and its guidance is at cursor.com/blog/claude-code-extension-security-notice. Cursor warns that uninstalling the extension alone may not be enough.
  • Rotate anything you shared. Cursor says some credentials users had sent to its support team were accessed. If you ever sent one, rotate it, and check provider logs for odd use.
  • Use Privacy Mode if your work is sensitive. Cursor says Privacy Mode users were not reported as affected.
  • Keep secrets out of prompts. Use environment variables and secret managers, and scope keys so a leaked one can do little. Set spend limits on any AI API key.
  • Audit vendor keys. Any key that lets a third party read your storage should be read-only, narrow and short-lived, with alerts on unusual access.
  • Watch for odd SSH keys. On a machine that may have had the bad extension, review authorized_keys for entries you did not add.
Pro Tip

Treat AI coding agent logs like a password manager export. If a tool stores your sessions, assume anything visible in a session could one day be read by someone else, and keep real secrets out of them.

A broader habit helps here. LumiChats is built so you can use several AI models in one place, and the general principle applies to any AI service you rely on: keep secrets out of prompts, use scoped credentials where an integration needs one, and avoid depending on one vendor for everything.

Frequently Asked Questions
01Was Anthropic hacked?

No. Both BleepingComputer and The Hacker News say Anthropic's legitimate extension was not compromised. Attackers placed tampered copies in Cursor's marketplace cache.

02Which Claude Code extension versions were malicious?

Versions 2.1.312 and 2.1.313, available for about 11 hours on October 29, 2026, according to Cursor's report as relayed by BleepingComputer and The Hacker News.

03Did attackers get my source code?

Cursor says customer code repositories were not affected, and Privacy Mode users were not reported as affected. Archived non-production Cursor repositories were accessed, and some credentials users sent to support were exposed.

04Did the backdoor work?

BleepingComputer reports researchers found coding errors that likely stopped it on many Windows and macOS systems without SSH enabled. Cursor still says up to 5,000 users could have received the build.

05Is Fireworks AI confirmed as the source?

It is Cursor's account that Fireworks exposed the key. We did not find a direct Fireworks statement about the Cursor key, though Fireworks separately disclosed its own October breaches.

Was this article helpful?

Found this useful? Share it with someone who needs it.

Free to get started

Claude, GPT-5.4, Gemini —
all in one place.

Switch between 40+ AI models in a single conversation. No juggling tabs, no separate subscriptions. Pay only for what you use.

Start for free No credit card needed
Aditya Kumar Jha
Written by
Aditya Kumar JhaLinkedIn

Published author of six books and founder of LumiChats. Writes about AI tools, model comparisons, and how AI is reshaping work and education.

Keep reading

More guides for AI-powered students.