When Fireworks AI first disclosed a security incident on Oct. 9, the damage it described was fairly contained: the company said the only data it knew had been accessed were secrets stored in evaluation jobs on its platform. Five weeks later the incident is much bigger. Fireworks has widened its own impact statement. Cursor, the AI coding editor, says an access key it gave Fireworks was stolen and used in a chain of intrusions that ended with a trojanized developer tool being served to Cursor users. And Uber has filed breach notices with several U.S. state attorneys general over driver's license data held by a vendor that TechCrunch identified as Fireworks.
Taken together, the disclosures show how far a compromise can travel when a vendor sits in the middle of other companies' data and holds credentials to their systems.
What Fireworks Says Happened
Fireworks AI hosts and serves AI models for other companies. Its updated bulletin (the bulletin index was last updated Nov. 9) says the impact now goes beyond evaluation-job secrets. The company has found exposure of account and billing information, including billing and shipping addresses, invoice and payment history, account balance and credit usage. Fireworks says it does not store payment card numbers. For a limited set of accounts, the attacker also reached secrets, datasets and incomplete models (a subset of incomplete model weights).
Fireworks says the evidence indicates links to a financially motivated cybercriminal group engaged in "LLMJacking," which it describes as selling tokens from unauthorized access to large language models to illicit routing gateways. Fireworks does not name a group, and no source attributes the intrusion to any named gang or nation-state.
The bulletin also describes a second phase. A third-party cloud provider that hosts Fireworks' research and development machines had its own incident on Oct. 23. On Oct. 24 the actor reached Fireworks development machines and ran an unauthorized workload there, and Fireworks says it detected that activity the same day. The company has notified law enforcement and is working with Palo Alto Networks' Unit 42 on forensic analysis. Its hardening steps so far: revoking credentials, restricting network pathways, tightening storage access, and turning off self-serve dataset upload and download.
How the Breach Reached Cursor
Cursor's account fills in what happened downstream. According to the company, the attacker got hold of an access key Cursor had issued to Fireworks, which Cursor says Fireworks exposed. With that key, the attacker copied data from a read-only storage bucket of coding-agent traces. Traces are logs of the prompts, code and tool calls produced when Cursor's AI agent works on a user's code. Cursor says the traces were limited, mostly from 2025, and came only from sessions that did not have Privacy Mode turned on. Privacy Mode is a Cursor setting that opts users out of having their data used for training, as The Hacker News explains.
Developers paste things into coding sessions that they shouldn't, and Cursor confirms that a small number of the traces contained unredacted credentials. The company says fewer than 0.01% of its customers had API credentials exposed this way.
The intrusion kept going. Cursor says the attacker got into select software-as-a-service tools it uses and archived non-production source code repositories. Then, on Oct. 29, the attacker used a stolen credential to put malicious copies of Anthropic's Claude Code extension on the cache that serves Cursor's extension marketplace.
Cursor is explicit about what this was not. It was not a vulnerability in Anthropic's software. Open VSX, the open extension registry, never carried the malware. The tampered packages lived only on Cursor's caching layer, and the company estimates that at most 5,000 users may have received one during a window of about 11 hours. The malicious versions are Anthropic.claude-code 2.1.312 and 2.1.313.
The Hacker News reports that the trojanized extension was pushed to users as an ordinary update and tried to install a persistent SSH backdoor, meaning a way for the attacker to get back into an infected machine later. That is why Cursor warns that uninstalling the extension may not be enough by itself.
Cursor says it published a security notice with indicators of compromise and remediation steps on Oct. 30. It says the attacker's last activity was on Nov. 1, that this activity was blocked, and that it believes the incident is contained.
Uber's Driver Data
The third company in the story is Uber. State filings, reported by TechCrunch, say Uber learned on Oct. 9 that hackers had stolen internal credentials used to access Uber data at a vendor. TechCrunch identified that vendor as Fireworks AI, and an Uber spokesperson confirmed the vendor incident. The California Attorney General's breach portal lists Uber Technologies, Inc. with a breach date of Oct. 7, 2026, and posts a multistate template notice. The California entry does not say how many California residents are affected.
The stolen data includes driver's licenses. According to TechCrunch, the filings count 2,567 affected people in Massachusetts, 3,430 in South Carolina and 1,873 in Rhode Island. The Illinois figure is still pending, and California drivers have also been notified. Uber has not given a total, and TechCrunch says the number is expected to rise. Fireworks confirmed to TechCrunch that there was unauthorized access affecting customer data but did not answer detailed questions.
The dates are worth reading carefully. Uber's Oct. 7 breach date and Oct. 9 notification come before the Oct. 23 cloud-provider incident and the Oct. 24 development-machine intrusion that Fireworks describes. So the cloud-provider episode is a later phase, not the starting point. None of the sources explains how the attacker first got in.
Why a Vendor Breach Travels So Far
Cursor gave Fireworks a key to a bucket. Uber had data stored with Fireworks. That is a normal setup for an AI inference vendor. But it means the vendor's credential store holds access to many customers' environments. In Cursor's case, the attacker used credentials found in exposed traces to reach internal tools and finally the extension marketplace cache. In general, vendor-issued keys may not get the same rotation schedule, scoping or monitoring as a company's own credentials, and this incident shows the cost of that gap.
What It Means for Defenders
Cursor users who run the Claude Code extension: Check which version of Anthropic.claude-code you have, and which versions you had during the Oct. 29 window. If version 2.1.312 or 2.1.313 was ever installed through Cursor's marketplace, treat that machine as possibly compromised, not just the extension. Cursor warns that uninstalling alone may not be enough, and The Hacker News reports the payload tried to set up persistent SSH access. Follow Cursor's remediation guidance in its incident update. Review the machine for unauthorized SSH access, and rotate any credentials that were stored on or used from that workstation, including SSH keys, cloud tokens, Git credentials and API keys. If you can't verify the machine is clean, reimage it.
Cursor customers in general: If you ran agent sessions without Privacy Mode, especially in 2025, assume anything you pasted into those sessions could be in the stolen traces. Rotate any API keys, tokens or passwords that appeared in prompts or agent output, even if you are not among the fewer than 0.01% of customers Cursor says had API credentials exposed. Going forward, keep secrets out of AI coding sessions entirely.
Fireworks AI customers: Fireworks says account and billing information was affected and, for some accounts, secrets, datasets and incomplete models. Rotate every credential Fireworks holds for you: API keys, cloud keys stored in evaluation jobs, and any keys you issued to Fireworks for access to your own storage or systems. Check your cloud logs for unexpected activity from those keys. Because Fireworks links the activity to an LLMJacking group, look for unexpected model usage or compute charges. Self-serve dataset upload and download is currently disabled, so plan any data transfers around that.
Uber drivers: If you got a notice, your driver's license information may be in criminal hands. Watch for identity fraud and phishing that uses your license details. Consider the credit freezes and monitoring your notice or your state offers.
Everyone with AI vendors: Make a list of the keys you have issued to third-party AI providers. Scope each one to least privilege. Make sure you can rotate them quickly, and alert on their use from unusual locations. For CMMC and other compliance programs, vendor-held credentials belong in your supply-chain risk assessment, not just the vendor's SOC 2 report.
Sources
- Fireworks AI: Update on October 2026 security incident
- Cursor: An update on our recent security incident
- TechCrunch: Uber confirms hackers stole thousands of drivers' license information in data breach
- California Attorney General: Submitted Breach Notification Sample (Uber Technologies, Inc.)
- The Hacker News: Cursor Says Stolen Fireworks AI Key Led to Trojanized Claude Code Extension Incident