TechPastWeek logo
Security10 min read

A Breach at an AI Cloud Provider Reached Uber, Notion and Cursor

Fireworks AI's breach cascaded into Uber, Notion and Cursor, exposing driver's licence data, customer chat transcripts and a trojanized coding extension.

The Uber office buildings in San Francisco's Mission Bay, where the company manages data for its rides and driver accounts.
On this page

Key takeaways

  • Uber notified 149,346 drivers that names, addresses, dates of birth and driver's licence numbers were stolen through a breach at its AI vendor Fireworks AI.
  • Notion said about 1% of its customers had chat transcripts from its Notion AI Insights feature exposed, and Cursor said credentials for less than 0.01% of customers were taken.
  • Cursor also confirmed that a stolen key led to a trojanized Claude Code extension reaching up to 5,000 users for about 11 hours on 29 October.
  • Fireworks attributed the intrusion to a financially motivated group that sells access to large language models, a scheme known as LLMJacking.
  • Fireworks hired Palo Alto Networks Unit 42 and notified law enforcement; it says the affected data includes billing details, secrets, datasets and some incomplete model weights.

Uber has begun notifying 149,346 drivers that their names, addresses, dates of birth and driver's licence numbers were stolen in a breach at Fireworks AI, a platform that hosts open AI models for other companies. The disclosure, filed with state attorneys general in November, shows how quickly a vendor incident can become a customer breach.

Related reading: AI Agents Found a Zero-Day in MONAI Label and Rooted a University Server · China-Linked Hackers Exploited a vCenter Zero-Day Weeks Before the Patch.

Fireworks confirmed the breach in October and has since published two updates. The company says the attacker reached its development systems through a third-party cloud provider, then used that access to take customer billing information, secrets, datasets and, for a limited set of accounts, incomplete model weights. It has hired Palo Alto Networks' Unit 42 to investigate and says it notified law enforcement. The company's bulletin describes the sequence and the data categories.

The customer list makes the incident more than a single-company story. Notion, the productivity and note-taking service, said about 1% of its customers had chat transcripts exposed. Cursor, the AI coding platform, said credentials belonging to less than 0.01% of its customers were taken, and that a stolen key was later used to plant a trojanized version of Anthropic's Claude Code extension in its marketplace. Fireworks has said it has tens of thousands of customers, including legal-tech company Harvey.

What Fireworks disclosed

Fireworks published its first bulletin on 9 October. In its later update, the company said a third-party cloud provider that hosted its research and development machines suffered its own security incident on 23 October. On 24 October, the attacker used that foothold to reach Fireworks development machines and run an unauthorized workload, which led to additional customer data being exposed. Fireworks says it detected the activity the same day and closed the known critical vulnerabilities.

The data categories are broad. Fireworks listed resource metadata such as identifiers and creation times; billing information including addresses, invoice and payment history, account balances and credit usage; and, for some customers, secrets, datasets and incomplete model weights. Payment card numbers were not stored by Fireworks and were not affected, the company said.

Fireworks also drew a line around its main products. It said it found no impact to its Inference service, its Training SDK or its Nexus product, all of which it describes as zero-data-retention by default. That distinction matters: the breach hit the internal systems Fireworks uses to run its business, not the inference paths most customers rely on for production traffic.

The company's remediation list is the standard one for a serious incident. It revoked and replaced credentials and service accounts, removed elevated privileges, restricted network pathways, tightened access to storage and analytics systems, and rebuilt logging and monitoring. It also disabled self-serve dataset uploads and downloads for batch inference and dataset-based training while the investigation continues.

Fireworks has not published a full list of affected customers or a final count. The figures from Uber, Notion and Cursor are their own notifications, and the investigation is ongoing. The company says its findings may change as the review continues.

Uber's blind spot

Uber's notice is the clearest example of how a vendor breach can bypass a customer's defences. The company said it had alerts configured to detect unauthorized access to its Fireworks environment. Those alerts did not fire, because the attacker was using Fireworks' own internal credentials rather than Uber's.

That is the core problem with shared responsibility in cloud services. Uber could monitor its own accounts, keys and network traffic. It could not see a session that looked, to its systems, like the vendor using its own access. The attacker did not need to steal an Uber password; it needed to compromise the company that already had permission to touch Uber's data.

The stolen data is sensitive but limited. Uber said the notices cover names, addresses, dates of birth and driver's licence numbers. It did not say that Social Security numbers, bank details or trip records were taken. The company began notifying drivers after being told about the breach on 9 October, and the count of 149,346 comes from the notice filed with Oregon's attorney general.

This is Uber's second known security incident this year. In August, the company reported an incident at its Uber Freight trucking unit after an extortion group claimed to have stolen data. The two events are separate, but together they weaken the argument that a company of Uber's size has its supplier risk under control.

Notion's chat transcripts

Notion's exposure is smaller in percentage terms and more revealing about how AI features store data. The company said the breach affected around 1% of its customers. The stolen material was chat transcripts from Notion AI's Insights feature, which lets users ask questions across their workspace. Notion's advisory sets out the timeline.

Notion said the transcripts include any input to or output from the model that was written to the chat. Much of the data was unattributable, meaning Notion could not always tell which customer it came from. The company disconnected its systems from Fireworks on 3 November, deleted the data it had stored there, and stopped connectivity between the two.

It also revoked Notion API keys that were found in affected user messages on 11 November. That detail is a reminder that chat logs often contain secrets users paste into them: keys, tokens, connection strings and internal URLs. A breach of a model provider can therefore become a breach of everything customers typed into the model.

Notion disabled the Insights feature and stopped serving open-weight models through Fireworks. It told affected customers to rotate any credentials they had pasted into Notion AI. The company's own systems were not breached, but its customers' data was, and the distinction will matter less to the people rotating keys.

Cursor and the trojanized extension

Cursor's disclosure is the most complex. The company said an attacker obtained an access key that Cursor had issued to Fireworks. The key was exposed in the Fireworks breach, and the attacker used it to copy data from a read-only storage bucket that Cursor used to share data with the vendor. Cursor's incident update describes the access and the later marketplace attack.

That bucket contained coding-agent traces from sessions that were not in Privacy Mode, the setting that opts users out of having their data used for training. Most of the traces dated from 2025. Some contained credentials that had not been edited out, exposing API keys belonging to less than 0.01% of Cursor's customers.

The attacker then used those credentials to reach selected SaaS tools used by Cursor. According to Cursor, the access covered archived non-production source code repositories, some containing business contact information; limited customer support data and email conversations; product usage metrics and contact details; summaries of some coding-agent conversations; and credentials that users had sent to the support team.

The most serious step came on 29 October. Cursor said the attacker used a stolen credential from the exposed data to place trojanized versions of Anthropic's Claude Code extension into the cache that served its extension marketplace. The malicious builds, versions 2.1.312 and 2.1.313, were available for about 11 hours. Cursor estimates that at most 5,000 users who requested those specific cached versions may have received them.

The extension attempted to install a persistent SSH backdoor, according to The Hacker News. Cursor took the marketplace offline, removed the builds, notified affected users and worked with the cloud provider hosting the attacker's server to shut it down. It also moved all inference traffic off Fireworks, deactivated the key and cut off the vendor's access to its datasets. The company says it has detected no further malicious activity since 1 November.

The LLMJacking economy

Fireworks attributed the intrusion to a financially motivated cybercriminal group whose goal is to sell access to large language models. The scheme is known as LLMJacking: attackers steal credentials or compute, run models at someone else's expense, and resell the capacity through unofficial gateways.

The economics explain the target selection. AI inference is expensive, and a stolen key can be worth more than a stolen database if it gives access to a frontier model with a large quota. The attackers do not need to exfiltrate customer records to profit. They need valid credentials and a way to route traffic.

That changes the risk model for AI providers. A traditional SaaS breach is about data. An AI platform breach is about data and compute, and the compute can be monetised immediately. It also makes the provider's own internal credentials a high-value target, because those credentials can reach every customer environment the provider supports.

Google's Threat Intelligence Group and Mandiant published research on LLMJacking on 2 November, describing the motives and techniques behind the rise in stolen model access. Fireworks cited that research in its update, which suggests the company sees its incident as part of a broader pattern rather than an isolated attack.

The supply-chain problem

The Fireworks case is a test of how much customers can rely on vendor assurances. Uber had alerts. Notion had a swift disconnection. Cursor had a read-only bucket and a Privacy Mode. None of those controls prevented the breach, because the attacker operated inside the trust boundary the vendor had been granted.

The practical lesson is that vendor access is an attack surface, not a compliance checkbox. Companies that give an AI provider access to data should know which credentials the provider holds, what those credentials can reach, and whether the access is logged in a way the customer can see. Many contracts do not require that transparency, and many customers do not ask.

The second lesson is about retention. Cursor's traces dated mostly from 2025, and some still contained credentials. Notion's chat logs contained API keys users had pasted in. Fireworks stored datasets and incomplete model weights for customers. Every one of those data stores increased the impact of the breach, and every one of them existed for a business reason that seemed reasonable at the time.

What customers should do

The immediate steps are familiar. Rotate any key, token or password that was shared with Fireworks or pasted into a tool that used it. Check logs for access from unfamiliar locations or at unusual times. Review which third-party applications can reach production data, and remove the ones that no longer need it.

The harder steps are structural. Give vendors scoped credentials rather than broad ones, and prefer short-lived tokens over static keys. Require vendors to notify customers quickly and to name the data categories affected. Keep an inventory of which AI features store customer content, and for how long. And test the assumption that a vendor's internal compromise will look like an external attack in your own logs. A vendor's own credentials often produce traffic that looks normal.

What to watch

Fireworks says its investigation is continuing, and the affected customer count may grow. The regulatory picture is also unresolved: state attorneys general are collecting notices, and Congress has shown interest in AI supply-chain security after a year of high-profile incidents. Whether the breach produces new disclosure requirements or remains a case study depends on what the investigation finds.

For Uber, the next step is notifications and identity protection for 149,346 drivers. For Fireworks, it is the rest of the investigation and the customers it has not yet named. For everyone else, the AI vendor with access to your data is part of your attack surface.

Cover: Uber's offices in Mission Bay, San Francisco. Photo by HaeB, CC BY-SA 4.0, via Wikimedia Commons.

Tags

fireworks-aidata-breachsupply-chainllmjackinguber

Frequently asked questions

What happened at Fireworks AI?
Fireworks AI, a platform for running open AI models, said an attacker compromised a third-party cloud provider that hosted its research machines, then reached Fireworks development systems on 24 October 2026. The intruder accessed customer billing data, secrets, datasets and some incomplete model weights.
How many Uber drivers were affected?
Uber notified 149,346 people, according to a notice filed with Oregon's attorney general. The stolen data included names, addresses, dates of birth and driver's licence numbers.
Was Notion breached directly?
No. Notion said the data was taken from Fireworks, a third-party AI vendor. About 1% of Notion customers had chat transcripts from the Notion AI Insights feature exposed.
What is LLMJacking?
LLMJacking is the theft of credentials or compute that lets an attacker run large language models at someone else's expense, often to resell access through unofficial gateways. Fireworks said the group behind its breach was financially motivated and focused on selling model access.
Share

Found this useful? Share it with your team.

TPW Editorial

TPW Editorial

Editorial Team

TechPastWeek's editorial team reports and edits every story we publish. We research each piece against primary and credible secondary sources, and a human editor reviews and signs off before it goes live.

Keep reading