Fireworks AI Breach Widens to Expose Billing Data, Model Weights

Fireworks AI's breach grew to expose billing data and model weights, hitting customers like Notion. Here's what Indian AI startups should take from it.

Nov 17, 2026 - 19:06
4 min read
 0
Fireworks AI Breach Widens to Expose Billing Data, Model Weights

Fireworks AI told its customers twice in three weeks that attackers had gotten into its systems. Each time, the list of what was actually taken got longer.

What happened, and when

Fireworks AI is a venture-backed company, last valued around $17.5 billion, that runs the behind-the-scenes infrastructure many other apps use to serve AI models cheaply and fast. Think of it as a wholesaler: your favourite AI-powered app might not run its own models at all, it might just rent compute and model access from a company like Fireworks.

In late October, an unauthorized party got into a third-party cloud provider that hosted some of Fireworks' research and development machines. The company's first disclosure said attackers had grabbed environment-variable credentials, the kind of stored passwords and API keys that let one system talk to another, for a limited number of accounts.

Then, on November 9, Fireworks widened that disclosure considerably. The attackers had also reached billing data: addresses, invoice and payment history, account balances, and credit usage. For a smaller group of customers, it went further still, pulling in secrets, datasets, and what the company called "incomplete model weights" — essentially partial copies of the trained AI models themselves.

The zero-retention promise, tested

Fireworks has long marketed a "zero data retention" option, where customer prompts and outputs are supposedly never stored after a request is processed. In its bulletin, the company was careful to wall that promise off from the breach:

"We have not detected any impact to Inference, Training SDK and Nexus, which are all zero data retention (ZDR) by default."

But the episode got messier once Notion, which uses Fireworks for parts of its AI features, weighed in. Notion has said it had a contractual zero-retention commitment from Fireworks, yet its own investigation reportedly found that some data had in fact been retained on the endpoints it used — the opposite of what it had been promised. Notion says the exposure affects fewer than 1% of its workspaces, and Cursor, the coding tool, was also pulled into the disclosure as another downstream customer.

Why one vendor's bad month is everyone's problem

The uncomfortable part of this story isn't really about Fireworks specifically. It's about how much of the AI boom now runs on a thin layer of infrastructure companies that most end users have never heard of. A startup building a chatbot or an AI writing tool rarely trains its own model from scratch — it rents inference from a provider like Fireworks, Together AI, or Replicate, stitches it into a product, and ships. That's smart economics. It also means a single breach two layers down the stack can quietly touch dozens of apps that look, on the surface, completely unrelated to each other.

  • Check whether any vendor holding your API keys or model weights has had a security incident in the last 12 months, not just whether they advertise "zero retention."
  • Ask vendors for proof of retention practices, not just policy language — Notion's experience shows the two don't always match.
  • Rotate API keys and secrets stored with any affected provider, even if your account wasn't named specifically.
  • Map which of your own customer-facing features actually depend on a third-party inference provider, so you know who to notify if something goes wrong upstream.

What it means for Indian companies building on rented AI

This matters directly for India's fast-growing crop of GenAI startups, many of which lean on exactly this kind of third-party inference layer to keep costs down while they scale. If an Indian company counts as a "data fiduciary" under the Digital Personal Data Protection (DPDP) Act — meaning it decides how and why personal data is processed — a breach at an overseas vendor it has plugged into doesn't let it off the hook. The Act puts the responsibility for notifying the Data Protection Board of India and affected users on the fiduciary, regardless of whose server the leak actually happened on. For a Bengaluru or Gurugram startup that built a product on top of a foreign model-serving platform, that's a reminder that vendor contracts and actual security practice need to match, because regulators and users alike will come looking at the company whose name is on the app, not the subcontractor three layers down.

The takeaway

Fireworks says its investigation is still ongoing and that it has notified law enforcement. That's the right move, but it doesn't undo the core lesson here: a marketing promise like "zero data retention" is only as good as the audit behind it. Every company renting AI infrastructure, in India or anywhere else, is now effectively trusting a vendor's internal controls with its own customers' data and its own product's reputation. This breach is a reminder to actually verify that trust rather than assume it.

Short URL: https://code24.in/0e86487c

What's Your Reaction?

Like Like 0
Dislike Dislike 0
Love Love 0
Funny Funny 0
Angry Angry 0
Sad Sad 0
Wow Wow 0
Ashif Sadique As an full-stack developer, I'm passionate about sharing tutorials and tips that aid other programmers. With expertise in PHP, Python, Laravel, Angular, Vue, Node, Javascript, JQuery, MySql, Codeigniter, and Bootstrap. To me, consistency and hard work are the keys to success.