The Fireworks AI breach is an AI supply chain problem, not a single-vendor story. Fireworks AI, a company that hosts and runs AI models for other software providers, has disclosed two security incidents within weeks, and reports say attackers reached customer datasets, model weights and Notion AI chat transcripts. If your teams use Notion AI, Cursor or similar tools, your data may have passed through Fireworks even though you never signed a contract with it.
This article explains what has been reported, who is reported as affected, how exposure happens through vendors you do not see, and what to check this week. It sticks to what the cited reporting says. It does not speculate on who carried out the attacks or which other companies may be involved.
What happened in the Fireworks AI breach
What Fireworks AI does
Fireworks AI is an inference provider. Inference is the step where a trained AI model takes a request, processes it and returns an answer. Many software companies do not run their own models. When you click an AI button inside a product, the product often sends your prompt, plus relevant documents or page content, to a provider like Fireworks. The provider runs the model on its infrastructure and sends the result back.
That arrangement is efficient, but it means the inference provider handles whatever your employees typed, pasted or uploaded. Some customers also use these providers to fine-tune models on their own data, which produces model weights: the learned parameters of a model. Weights tuned on private data can represent real intellectual property.
Two incidents within weeks
Fireworks published a security bulletin about an October 2026 incident and later posted an update describing further activity. Bloomberg reported that the Nvidia-backed company notified customers of its second hack in weeks. The Hacker News reported that Fireworks confirmed a wider breach affecting customer datasets, model weights and Notion AI chats.
The pattern matters as much as the details. The first disclosure was not the full picture. When a vendor is still investigating, the scope can grow, and your own assessment should stay open until the vendor's investigation closes.
Which companies and products are reported as affected
The reporting so far names the following. Treat this as a starting point, not a complete list.
| Company or product | What was reported | |---|---| | Notion AI | Notion AI chats were affected, per The Hacker News | | Cursor | A data breach at an AI provider led to malware at Cursor, per Techzine | | Uber | Uber drivers were affected by a data breach at a third-party provider, per SC Media | | Fireworks direct customers | Customer datasets and model weights were affected |
The Notion AI chat transcripts breach is the one most leadership teams will feel first. Notion workspaces commonly hold meeting notes, plans, hiring discussions and client material. If employees used Notion AI to summarize or draft from those pages, that content may have been part of what was sent for processing.
The Cursor side of the Fireworks AI breach matters for any company with software developers. Cursor is an AI coding tool. Developer environments tend to contain source code, configuration files and credentials, so anything touching them deserves careful follow-up.
How you can be exposed without a Fireworks contract
Most companies affected by an AI vendor breach like this one never chose the AI vendor. The chain usually looks like this:
- Your company contracts with a software provider, such as a notes, coding or customer support tool.
- That provider adds AI features and routes AI requests to an inference provider.
- The inference provider is listed as a subprocessor: a third party that handles customer data on the software provider's behalf.
Your contract is with step one. Your data sits at step three. This is the core of AI supply chain risk.
Several things make it easy to miss:
- Subprocessor lists change quietly. Vendors often update a web page or send an email to an admin address that nobody monitors.
- AI features are switched on by default in many products, sometimes after you signed the contract.
- Shadow AI bypasses the admin entirely. When employees use AI tools on personal or free accounts, no one at your company receives the breach notice. Our article on building a shadow AI policy covers how to bring that use into view.
- Prompts carry unstructured data. A pasted contract or a credential in a chat window does not show up in any data map.
What to do this week about the Fireworks AI breach
These steps are ordered so the fastest, highest-value checks come first. Most can be done by existing IT and operations staff with an executive sponsor.
- Find the notices. Check admin and security inboxes for messages from Notion, Cursor and any other AI-enabled vendor. Review each vendor's trust or status page. If you use these tools and have heard nothing, ask the vendor directly in writing.
- List your AI features and their subprocessors. Start with the ten software tools that hold your most sensitive data. For each, record which AI features are enabled and which subprocessors handle AI requests. Look for Fireworks AI by name, but write down every inference and model hosting provider you find. You will need that list again.
- Rotate exposed credentials. Identify API keys, access tokens and passwords that may have been pasted into AI chats, stored in AI-enabled workspaces, or present on developer machines using affected tools. Rotate them, starting with cloud, source code, payment and production database access. Then review logs for unusual use of those keys in recent weeks.
- Assess what data employees put into AI features. Ask team leads a simple question: what kinds of information did your team use with Notion AI or similar features? Customer personal data, health information, contracts, financials, deal material and credentials are the categories that drive notification decisions.
- Follow vendor-specific guidance. If your engineers use Cursor, ask Cursor what steps it recommends for checking developer machines and follow any guidance it provides. Do not improvise around a vendor's incident guidance.
- Write down what you did. Record what you checked, when, who checked it and what you found. Counsel, your insurer, customers and your board may all ask, and a dated record is far more credible than recollection.
Questions to ask your AI and SaaS vendors
Third-party AI provider security is now part of every vendor conversation. Send these questions to each vendor whose AI features touch sensitive data:
- Which third parties host or run the models behind your AI features?
- Was Fireworks AI among them, and was our data processed there during the affected period?
- What exactly do you send to inference providers: only the prompt, or also documents, page content, files and metadata?
- How long do inference providers retain prompts, outputs and uploaded files? Is any of it used to train models?
- How and when will you notify us if a subprocessor has a security incident?
- Do your contracts with subprocessors carry the same security and confidentiality terms you agreed with us?
- Can we disable AI features, limit them to certain workspaces, or choose which provider processes our data?
Vague answers are themselves useful information. A vendor that cannot name its inference providers cannot tell you whether you were affected.
Notification and regulatory obligations
Whether you have to notify anyone depends on what data was involved, whose data it was, where those people live, your sector and your contracts. Your vendor contracts usually say when and how the vendor must tell you about an incident. Your own obligations generally begin when you learn your customers' or employees' data may have been affected.
The main sources of obligation to review with counsel:
- State breach notification laws. Virginia and the District of Columbia each have their own statutes, as do other states. They generally turn on specific categories of personal information.
- Sector rules. Health data, financial data and government contract data each carry their own reporting regimes. The FTC, for example, has updated its Health Breach Notification Rule with new provisions to protect users of health apps and devices.
- Your own privacy statements. The same FTC page notes that the agency routinely evaluates the privacy representations a company makes against its data handling practices. If you told customers their data stays with you, check that against where your AI features send it.
- Customer contracts. Many enterprise and government customer agreements require notice of suspected incidents involving their data, often on tighter timelines than the law. These are frequently the obligations that bite first.
The practical step this week is to identify which of these could apply, based on the data assessment in step four above, and bring in legal counsel before you draw conclusions.
Building AI vendor risk into governance, contracts and incident response
The immediate checks will settle this incident. The lasting fix is making AI vendors visible and governed as a routine matter.
Keep an AI inventory that includes subprocessors
An AI inventory that lists only the tools you bought misses the point of this breach. For each tool, record the business owner, AI features enabled, data types allowed, subprocessors involved in AI processing, key contract terms and the date of the last review. Assign someone to watch for subprocessor changes. This inventory belongs at the center of your AI governance and AI policy work, and it connects directly to the broader view that vendor risk is the control plane, not a checklist.
Tighten contract terms at renewal
When AI-enabled vendors come up for renewal, ask for:
- Advance notice of new AI subprocessors and a right to object.
- Flow-down of your security and confidentiality terms to those subprocessors.
- Clear incident notice requirements that cover incidents at subprocessors, not only at the vendor.
- Limits on retention of prompts and outputs, and a commitment not to train on your data without consent.
- Cooperation in investigations, including access to the facts you need for your own notification decisions.
Add the vendor-of-a-vendor scenario to incident response
Most incident response plans assume the incident starts inside your network or at a direct vendor. Add a scenario where a provider two steps removed is breached. Decide in advance who receives and triages vendor notices, who leads the data assessment, who calls counsel and who briefs leadership. Run it as a short tabletop exercise. If you want an outside view of where these gaps sit, a cybersecurity risk assessment (expert-led) is one place to start.
Set rules for what goes into AI features
Governance is not only about vendors. Give employees a short, plain rule set: which tools are approved, which data types may never be entered into AI features, and what to do if they have already pasted something sensitive. Credentials should never go into a prompt.
The lesson to carry forward
The Fireworks AI breach shows that your exposure to AI is defined by the full chain of providers behind your tools, not by the contracts you signed. This week, find the notices, map your AI subprocessors, rotate credentials that may have touched AI tools, and understand what data your people put into AI features. Over the next quarter, put subprocessors into your AI inventory, your contracts and your incident response plan. Companies that can answer the question of where their AI data goes, quickly and with evidence, will handle the next AI vendor incident calmly.
If you want help assessing your exposure, start with the free AI risk assessment. It takes five minutes and gives you a baseline score right away.
For the field guide version of what I publish here each week, pick up a copy of Exposed: Inside Risks and the New Architecture of AI Defense on Amazon.
NightFortress works with executives, founders and leadership teams in Northern Virginia and the DC metro area to assess exposure, govern risk, and build security programs that match the actual threat landscape. Contact us to start a conversation.
The information in this article is for educational and informational purposes only. It is not intended as legal, compliance, or professional cybersecurity advice for any specific organization. Consult qualified professionals before making security or compliance decisions.

