Most companies now use AI through a chain of vendors. You buy a product with an AI feature, that product calls a model provider, and the model provider may run on yet another company's infrastructure. The Fireworks AI breach, which became public in stages through October and November 2026, shows what happens when one link in that chain is breached.

What happened

Fireworks AI is an inference provider: it hosts and runs AI models, including open-weight models, for other companies. It first disclosed unauthorised access to its systems in October 2026. In later updates it widened that disclosure to cover account and billing information and, for a limited set of accounts, secrets and datasets. Fireworks has since said it was dealing with two separate security incidents, not one.

The effects reached at least three well-known customers:

  • Notion said the breach exposed Notion AI chat transcripts from fewer than 1% of its workspaces. The transcripts had been stored at Fireworks for Notion's Insights feature, which analyses stored chats, even though Notion had a contractual zero data retention commitment from Fireworks.
  • Uber is notifying about 149,000 drivers that information including driver's licence details was taken. TechCrunch reported on 16 November that Uber said the notices relate to a security incident at a third-party vendor, Fireworks.
  • Cursor said an attacker obtained an access key Cursor had issued to Fireworks, and that this led to a trojanised Claude Code extension appearing on Cursor's marketplace.

Investigations are still running, and some dates and details differ between reports. The pattern, though, is already clear: one AI supplier was breached, and the damage showed up at its customers and at their users.

What it means for leadership teams at mid-size companies

You probably do not buy from Fireworks directly. That is the point. Few of Notion's or Uber's end users knew Fireworks was involved at all. If your teams use AI features inside SaaS tools, or your engineers call model APIs, you have suppliers like this too, and you may not have a list of them.

Three things in this story apply to almost every company using AI.

Zero data retention is narrower than it sounds. A zero data retention clause usually covers specific endpoints or ways of using a service. Features built on top, such as analytics, chat history, evaluation or batch jobs, can store data under different terms. Notion's experience shows a company can have the clause in its contract and still have data sitting at the vendor.

Keys you give vendors are part of your attack surface. By Cursor's account, its incident started with a key it had issued to a supplier. Any API key, token or service account you hand to an AI vendor is only as safe as that vendor.

Your customers will hold you responsible. Uber sent the notices to its drivers. Notion explained the exposure to its workspaces. When an AI supplier is breached, the company whose name the customer knows does the explaining, whatever the contract says.

What to do this quarter

None of this needs a large programme. A mid-size company can cover the basics in a few weeks.

  1. Map your AI vendor chain. List every AI model provider you call directly, every SaaS tool where staff use AI features, and the sub-processors each of them names for AI. Ask vendors who runs inference for them if they do not say.
  2. Ask what is actually stored. For each vendor, find out which features keep prompts, outputs, files or chat history, for how long, and where. Check whether your zero data retention terms cover those features or only basic model calls.
  3. Turn off what you do not use. Disable chat history, analytics and admin insight features that store AI conversations if nobody relies on them. Data that was never kept cannot leak.
  4. Scope and rotate the keys you hand out. Give each vendor its own key with the narrowest permissions it needs, set expiry dates, and know how to revoke every key quickly. Log what each key does.
  5. Keep sensitive data out of prompts by default. In RAG and agent systems, retrieve only what the task needs, mask personal data where you can, and keep the source documents in systems you control.
  6. Write the incident path down. Agree who decides, who tells customers and how fast, if an AI supplier reports a breach. Check that your vendor contracts require prompt notice and say what data was affected.
  7. Put AI suppliers in your normal vendor review. The same security questionnaire and annual review you use for payroll or CRM vendors should apply to model providers and AI features.

Where I can help

This kind of review is a normal part of my Fractional AI Lead and Advisory service, where I review architecture and vendors with your team on a monthly retainer. If you are earlier on and still deciding where AI fits, an AI readiness assessment covers data handling and vendor risk alongside use cases. And if you are building retrieval systems, my production RAG work starts with the question of what data should leave your systems at all.

Sources