Most people have never heard of Fireworks AI, and that is sort of the point. It is one of several companies that run AI models in the background for other businesses, the kind of vendor that shows up nowhere in a product's marketing but ends up processing real customer data anyway. In October, attackers broke into Fireworks' systems not once but twice, and the fallout now includes stolen Uber driver license numbers and leaked Notion AI chat transcripts, two completely different products whose only connection is that they both quietly relied on the same backend vendor.
How the attackers actually got in
According to Fireworks' own disclosure, the intrusion did not start with Fireworks at all. On October 23, a third-party cloud provider that hosted some of Fireworks' research and development machines was compromised first. The next day, the attacker used that foothold to reach a Fireworks development machine, ran an unauthorized workload, and escalated permissions through several hops until it could act as a cloud service account, essentially working its way up from an outside contractor's weak link to something close to internal access. Security researchers have tied the activity to LLMjacking, where stolen access to AI infrastructure gets resold on criminal marketplaces rather than used directly by the people who stole it.
Fireworks disclosed a first, narrower version of the incident in early October, describing exposed API keys and account secrets. By November 9, the company widened that disclosure considerably, confirming that billing information, datasets, and incomplete model weights had also been accessed for a limited set of accounts, and that tens of thousands of files in some cases contained customer data.
The Uber piece: real driver's license numbers
Uber confirmed that hackers stole personal information belonging to thousands of its drivers through the Fireworks breach, specifically driver's license details. Uber's notifications named 2,567 affected people in Massachusetts, 3,430 in South Carolina, and 1,873 in Rhode Island, close to 7,900 drivers across just those three states. That is not an email address or a billing record. A driver's license number tied to a name is the kind of data that enables actual identity theft, not just spam.
The Notion piece: AI chats that were supposed to disappear
Notion's exposure is a different flavor of damage, and arguably the more uncomfortable one to explain to users. Notion uses an AI feature called Insights, and some of the conversation data behind it was processed through Fireworks. Notion says it had a contractual commitment from Fireworks for zero data retention on that specific endpoint, meaning the chat content was supposed to be processed and then gone, not stored anywhere an attacker could later find it.
That is not what happened. The chat transcripts ended up sitting in a dataset used for a classification job, and that dataset is exactly what got swept up in the breach. Notion says fewer than 1% of its workspaces were affected and that the vast majority of the exposed transcripts could not be tied back to a specific user or workspace, which limits the damage but does not erase the underlying problem: a "zero retention" promise applied at the model level, while a completely separate storage layer downstream kept a copy anyway.
Why this is worse than a typical single-company breach
A breach at one company is bad enough. A breach at an AI infrastructure vendor is worse by design, because that one vendor sits underneath many unrelated products at once. Fireworks has said it works with more than 10,000 companies, and its publicly listed clients include well-known names across retail and tech. Uber and Notion are the two that have confirmed actual data exposure so far, but the structural risk applies to every customer using that same backend, whether or not their data happened to be in the specific files an attacker reached.
This is the pattern security researchers have been warning about as more companies bolt AI features onto existing products: the AI feature itself might be well designed, with sensible retention rules and encryption, and still get compromised through a vendor three layers removed from anything the end user ever sees or agrees to.
What this means if you use Notion or drive for Uber
- If you are an affected Uber driver, Uber is notifying you directly. Treat that notice as real and follow the credit monitoring or identity protection steps it offers, since a stolen driver's license number is reusable fraud material for years, not just a one-time risk.
- If you use Notion AI features, there is no indication regular account passwords or files were exposed, the issue is specifically tied to AI chat transcripts processed through the Insights feature. Notion has disabled the affected integration and deleted its data from Fireworks' systems, but it is still reasonable to avoid pasting anything sensitive, like credentials or financial details, into an AI chat box inside any product, since you often cannot verify every vendor behind the scenes.
- More generally, "zero data retention" is a contract term, not a physical law. This incident shows a company can sincerely commit to it at one layer of its system while a different team, a different pipeline, or a different dataset quietly keeps a copy anyway.
The takeaway
The uncomfortable lesson here is not really about Fireworks, Uber, or Notion specifically. It is that the AI layer increasingly sitting underneath ordinary apps is itself a new, often invisible point of failure, run by companies most users will never interact with directly or even know exist. Reading a privacy policy from the app you use does not tell you much about the five vendors processing data behind it. Until that changes, the only real defense is to assume that anything typed into an AI feature, inside any product, could eventually end up somewhere you did not expect.