Cursor Ties Claude Code Trojan to Stolen Fireworks AI Key

Cursor Ties Claude Code Trojan to Stolen Fireworks AI Key

Estimated reading time: 7 minutes · Last updated:

Cursor says the trojanized Claude Code extension that reached its marketplace on 29 October 2026 traces back to one stolen access key. The AI coding company had issued that key to Fireworks AI, its inference provider, and Fireworks exposed it during a breach. An attacker used the key to copy coding-agent traces that still held user credentials. Those credentials led into Cursor's SaaS tools. From there, the attacker planted a malicious build of Anthropic's extension in the marketplace cache. Cursor caps the exposure at 5,000 users over roughly 11 hours. The same Fireworks AI breach has since reached Uber and Notion, which makes a single supplier the common thread across all three companies.

At most 5,000 users who requested specific cached versions of the extension on October 29 potentially received the malicious build.

Cursor

Key takeaways

  • Root cause: Cursor says an attacker took an access key it had issued to Fireworks AI, which Fireworks exposed, and used it to copy data from a shared read-only storage bucket.
  • Credential spill: Unredacted credentials in Cursor coding-agent traces, mostly from 2025 sessions without Privacy Mode, exposed API credentials for less than 0.01% of Cursor customers.
  • Malicious build: Trojanized Claude Code versions 2.1.312 and 2.1.313 sat in Cursor's marketplace cache for about 11 hours on 29 October 2026 and reached at most 5,000 users.
  • Containment: Cursor has moved all inference traffic off Fireworks AI and reports no further malicious activity since 1 November 2026.
  • Wider fallout: Uber confirmed the theft of driver's licence data for thousands of drivers through a service provider that TechCrunch identified as Fireworks AI, and Notion has also confirmed data theft.

How a Key Issued to Fireworks AI Opened a Path Into Cursor

The chain began with a credential Cursor gave to a supplier. Cursor had issued an access key to Fireworks AI so the two companies could exchange data through a read-only storage bucket. In a Friday update on its blog, Cursor said Fireworks exposed that key. The attacker then used it to copy what the bucket held.

The bucket held a limited set of coding-agent traces, which are logs of what Cursor's AI agent did during a user's session. They came from sessions where Privacy Mode was switched off. Privacy Mode is a Cursor setting that lets users keep their data out of model training. Most of those sessions dated from 2025, and their data had been cleared for training.

Some traces still contained credentials that nobody had removed. Cursor says this exposed API credentials belonging to less than 0.01% of its customers. The attacker used them as a way into selected software-as-a-service tools the company relies on. Through those tools the intruder reached several kinds of data. These included archived non-production source code repositories, some of which held business contact details. They also included limited customer support records and email threads, product usage metrics and contact information, summaries of some agent conversations, and credentials that users had sent to Cursor's support staff.

The weak point was not Cursor's production systems. It was a copy of training data kept for a partner. Secrets pasted into agent sessions stayed in the traces, so a bucket meant for model work also served as a store of credentials.

What the Trojanized Claude Code Build Did on 29 October

The second phase came on 29 October 2026. The attacker took a credential from the data stolen out of Cursor's internal systems and used it to push trojanized copies of Anthropic's official Claude Code extension into the cache that serves Cursor's extension marketplace. To developers, the poisoned copy looked like a routine update.

The malicious build tried to install an SSH backdoor that would stay on the machine and give the attacker lasting remote access. Cursor says the build was live for about 11 hours before the company pulled it. Cursor then took the whole extension marketplace offline. It also got the cloud company running the attacker's server to shut that machine down.

Cursor says no more than 5,000 users could have received the bad build. Only users who requested particular cached versions on that day were at risk. According to Cursor, it has contacted everyone who pulled version 2.1.312 or 2.1.313 from the marketplace in that window and helped them revoke credentials that may have leaked. Cursor explained the root cause nearly two weeks after the extension was compromised.

How Cursor Says It Contained the Intrusion

Cursor found that the intruder had gone further than the shared bucket. It then deactivated the vendor's key, revoked and rotated credentials, and searched for signs of deeper access. It also switched off some internal and product services for a time, so some Cursor features were unavailable to users.

Cursor has now cut ties with Fireworks AI. It has taken all of its inference traffic off Fireworks, deactivated the key it had issued, and removed Fireworks' access to every data set the two companies had shared.

Cursor says it has detected no further malicious activity since 1 November 2026. It also says it has found no evidence that the incident reached several areas. These are customer code repositories, its production environment, the Cursor app or its official update channel, the company's other apps, other extensions and plugins in its marketplace, and anyone who had Privacy Mode turned on.

These conclusions come from Cursor's own investigation. They describe what the company has found so far. They do not guarantee that nothing else was touched.

Fireworks AI Breach Now Reaches Uber and Notion

Uber disclosed its exposure in a data breach filing with the state of California. The company said a service provider told it on 9 October 2026 that an outside party had stolen the provider's internal credentials and used them to log into Uber's account with that provider. By 13 October 2026, Uber's investigation had found that the intruder used that access to reach Uber data linked to the provider.

Uber's filing does not name the provider. TechCrunch reported that it is Fireworks AI. Uber has confirmed that the attackers took driver's licence information for thousands of drivers. It has not disclosed how many people were affected in total.

Notion, the online productivity startup, has also confirmed that its data was stolen in the Fireworks AI hack. That makes at least three companies whose data was exposed through this one supplier.

All three cases involve a supplier that held keys and credentials for its customers' systems. Neither the Cursor nor the Uber disclosure says how the attacker first got into Fireworks AI.

Companies reporting exposure linked to the Fireworks AI breach
Company How the attacker got in Data affected Key dates
Cursor Access key issued to Fireworks AI, then credentials found in agent traces Agent traces, support data, archived repos, a trojanized Claude Code build 29 October 2026 extension attack; quiet since 1 November 2026
Uber Service provider's internal credentials, provider named by TechCrunch as Fireworks AI Driver's licence information for thousands of drivers Notified 9 October 2026; access confirmed 13 October 2026
Notion Fireworks AI hack Company data, scope not detailed Not stated

Where the Fireworks AI fallout goes from here

The case for

  • Cursor reports no malicious activity since 1 November 2026, and the poisoned extension was live for about 11 hours.
  • Exposure from the trojanized build is capped at 5,000 users, and Cursor has contacted those who installed versions 2.1.312 or 2.1.313.
  • Cursor has removed Fireworks AI's access to shared data sets and moved its inference traffic elsewhere, closing the original entry point.

The case against

  • Uber has not said how many people its breach affected, so the full scale of the Fireworks AI incident is still unknown.
  • Credentials that users sent to Cursor support were among the stolen data, and they stay usable until their owners rotate them.
  • At least three companies have now been hit through one supplier, and the disclosures do not explain how Fireworks AI was first breached.

What to be careful about

  • Machines that installed version 2.1.312 or 2.1.313 may still carry the SSH backdoor if it was installed and has not been removed.
  • Credentials that users pasted into agent sessions or sent to Cursor support were part of the stolen data and could be reused.
  • Driver's licence data taken from Uber can be used for identity fraud against thousands of drivers.
  • Other Fireworks AI customers may not yet know whether their data or keys were exposed.

The bottom line

The Cursor incident shows how a breach at one AI supplier can spread to its customers. A key held by Fireworks AI gave access to a training-data bucket. Credentials left in agent traces opened SaaS tools. A stolen internal credential then let the attacker poison an extension cache that developers trusted. Cursor says the malicious Claude Code build reached at most 5,000 users and that it has seen no new activity since 1 November 2026. Uber and Notion are dealing with the same supplier breach, and Uber has yet to say how many people were affected. Anyone who installed version 2.1.312 or 2.1.313 that day should already have heard from Cursor. Credentials typed into agent sessions that fed training data were among the stolen data.

What to watch

  • Watch for Uber to state how many drivers and other users were affected by the Fireworks AI breach; no date has been given.
  • Watch whether more Fireworks AI customers beyond Cursor, Uber and Notion disclose exposure; no timetable has been set.
  • Watch for any further Cursor update on whether malicious activity has resumed after the quiet period it has reported since 1 November 2026; no date for that update has been announced.

Frequently asked questions

Did Cursor Privacy Mode protect users in the Fireworks AI breach?

Cursor says it has found no evidence that the incident affected users who had Privacy Mode turned on. The leaked traces came from sessions without Privacy Mode, most of them from 2025, whose data had been cleared for model training.

Which Claude Code extension versions in Cursor were malicious?

Versions 2.1.312 and 2.1.313, served from Cursor's marketplace cache on 29 October 2026, carried the trojan. They were available for about 11 hours, and Cursor says at most 5,000 users may have received them.

Was the Uber data breach linked to Fireworks AI?

Uber's California filing names only a service provider, which notified Uber on 9 October 2026. TechCrunch reported that the provider is Fireworks AI, the same company whose exposed key started the Cursor incident.



Share:

Categories

Newest course every month

Advertise your offline course to a wider audience with our landing page.

You May Also Like

Cursor says a key it issued to Fireworks AI was exposed in the Fireworks breach, giving an attacker the path...
United Airlines is restricting third-party AI agents from booking flights and managing reservations, citing credential leaks and federal check-in rules.
Arctic Wolf Labs has linked the AntiFox AI pentesting platform to intrusions on IIS web servers and AWS accounts that...