AI Security · Supply Chain Risk

AI Inference Provider Risk: Lessons from the Fireworks AI Breach

By Infonaligy · Updated November 13, 2026 · 8 min read

AI SUPPLY CHAIN // INFERENCE RISK

When you sign up for an AI feature, you are rarely trusting just one company. The application vendor often sends your prompts to an inference provider, and that provider runs on its own cloud hosts. The Fireworks AI breach, which reached Notion AI chat transcripts, shows how a security incident two or three links down that chain can expose data you shared with a vendor you trust. The fix is not to stop using AI. It is to map the chain, tighten contracts, minimize what you send and have a playbook ready for when a fourth party is breached.

What happened in the Fireworks AI breach?

Fireworks AI, an inference platform that hosts and serves AI models for other companies, has disclosed two related incidents. According to its original security bulletin, Fireworks identified unauthorized use of internal credentials on October 6, 2026, and found that environment variables in evaluation jobs, including API keys and tokens stored as secrets, were likely captured in a limited number of customer accounts.

Fireworks' updated notice, last updated November 9, adds a second event. A third-party cloud provider hosting Fireworks' R&D machines had a security incident on October 23, and on October 24 the unauthorized party used it to reach Fireworks development machines and run a workload. Fireworks says exposed data may include account and billing information (it says payment card numbers were not stored and were not affected), secrets for a limited set of accounts, datasets for certain customers and a subset of incomplete model weights. It engaged Palo Alto Networks Unit 42 for forensic analysis and links the activity to a financially motivated group that sells stolen model access to illicit routing gateways, a practice it calls LLMjacking. Fireworks also says it detected no impact on its zero data retention services, including inference. Coverage from The Record, BleepingComputer and The Register tracks the same timeline.

How did a Fireworks breach reach Notion customers?

Notion used Fireworks for its Insights feature and as an inference provider for open-weight models. In its customer notice, Notion says Notion AI chat transcripts, meaning the inputs and outputs of AI chats, may have been in data accessed in the incident, affecting fewer than 1% of workspaces. The Record reports that Notion learned of the incident on November 2 and that Fireworks told it on November 6 that customer data may have been accessed.

Notion says it disconnected from Fireworks, had its data in the Fireworks system deleted, disabled Insights and began auditing critical vendors for zero data retention compliance. It also told customers that anyone who pasted passwords, API keys or other credentials into Notion AI during the timeframe should rotate them right away. That last instruction is the lesson for every IT leader: data typed into an AI tool can travel further than the tool itself.

What is AI supply chain risk?

AI supply chain risk is the exposure created when your data passes through vendors you did not choose directly. A typical chain looks like this:

  • First party: Your organization and your users, typing prompts and uploading files.
  • Third party: The SaaS vendor whose AI feature you bought, such as a workspace, CRM or help desk tool.
  • Fourth party: The inference provider or model host the vendor sends prompts to, often listed only on a subprocessor page.
  • Fifth party: The cloud or GPU provider hosting that inference provider's machines, as in the October 23 event.

Each link holds credentials, logs, datasets or transcripts. Your security review usually covers only the second link.

How do you map your AI subprocessor chain?

Start with an inventory of every AI feature in use, then trace where each one sends data. A practical approach:

  • List AI features, not just AI vendors: Many SaaS tools you already own have switched on AI assistants. Check admin consoles for AI settings.
  • Pull subprocessor lists: Record each inference provider, model host and region named on the vendor's subprocessor page or trust center.
  • Classify the data flow: Note whether the feature sees chat text, documents, customer records or code, and whether it stores transcripts for analytics or training.
  • Subscribe to changes: Sign up for subprocessor change notices so a new inference provider does not appear silently.

Our AI vendor risk automation guide shows how agents can keep this register current instead of relying on a yearly questionnaire.

Which contract terms matter most with AI vendors?

Contracts decide what a fourth-party breach can expose and how fast you hear about it. Ask for these terms in writing:

TermWhat to ask for
Data retentionZero data retention for prompts and outputs, or a short, stated retention period
Training useNo use of your content to train or tune models without opt-in
Analytics useClarity on whether transcripts feed product analytics features
Breach notificationA defined notification window that covers incidents at subprocessors
Subprocessor changesAdvance notice and a right to object before a new inference provider is added
DeletionConfirmed deletion at subprocessors when the contract ends or a subprocessor is removed

Zero data retention options deserve special attention. Fireworks says its zero data retention services showed no detected impact, while stored secrets, datasets and incomplete model weights were affected; the Notion transcripts exposed were stored for an analytics feature. Data that is never stored cannot be stolen later.

How should you handle secrets and credentials in AI tools?

Treat every AI prompt box as a place secrets should never go. Both the Fireworks credential theft and Notion's rotation advice point to the same habit.

  • Policy: Make it explicit that passwords, API keys, tokens and connection strings never go into AI chats.
  • Secret scanning: Scan prompts at an AI gateway or browser extension, and scan repositories and wikis where people paste keys.
  • Vaulting: Keep keys in a secrets manager and inject them at runtime instead of storing them in job environment variables.
  • Rotation playbook: Know who owns each key, how to rotate it, which applications break and how to verify the old key no longer works. Fireworks reminds customers that deleting a stored secret does not invalidate it; only revoking it at the issuer does.

An AI gateway control plane is a practical place to enforce scanning, redaction and logging for every model call in one spot.

How do you reduce what reaches the inference provider?

Data minimization limits the blast radius of any breach downstream. Redact names, account numbers and personal details before inference where the task allows it. Send only the fields a feature needs, not whole records. Log which systems and users send data to which models, and keep those logs where your security team already watches. For guidance employees can follow, see our guide to keeping company data safe with AI.

For sensitive workloads such as legal, financial, health or source code, consider hosted or private inference, where models run in an environment you control, with your own retention, logging and endpoint protection. Our AI DevOps team builds the pipelines that keep those deployments patched and monitored.

The bottom line

Your AI vendors have their own vendors. Map the inference subprocessors behind every AI feature, require zero data retention, no training use and fast breach notice that covers subprocessors, keep secrets out of prompts and minimize what you send. Have a fourth-party incident playbook ready before you need it.

What should you do when a fourth-party AI provider is breached?

Move quickly and assume exposure until the vendor proves otherwise. A short incident sequence:

  • Confirm exposure: Check whether your vendors use the breached provider, using your subprocessor map and the vendor's notice.
  • Rotate credentials: Revoke and reissue any keys stored with the provider or pasted into affected AI tools, with the issuing service.
  • Review content: Use vendor review tools, as Notion offers, to see which transcripts or files were in scope, and look for sensitive data.
  • Hunt for misuse: Watch for unusual API usage and unexpected AI billing spikes, a common sign of LLMjacking.
  • Assess obligations: Involve legal on customer, regulatory and contractual notification duties.
  • Close the loop: Update the subprocessor register, contract terms and user guidance based on what you learned.

Our AI security practice helps teams build this playbook, and our AI consulting team can prioritize which vendors to review first.

Frequently asked questions

What is a fourth-party AI provider?

It is a vendor your vendor uses. When a SaaS tool adds an AI feature, it often sends prompts to a separate inference provider or model host. You have no direct contract with that provider, but your data can still reach it, which is why it belongs in your vendor risk review.

Was Notion data exposed in the Fireworks AI breach?

Notion says Notion AI chat transcripts sent to Fireworks may have been in data accessed in the incident, affecting fewer than 1% of workspaces. Notion disconnected from Fireworks, disabled its Insights feature and advised anyone who pasted credentials into Notion AI to rotate them.

Does zero data retention prevent this kind of exposure?

It reduces it. Data that an inference provider never stores cannot be taken from storage later. Fireworks says its zero data retention services showed no detected impact, while stored secrets, datasets and incomplete model weights were affected; the Notion transcripts exposed were stored for an analytics feature. Retention, analytics use and subprocessor terms all still matter.

What should we do first if an AI vendor's provider is breached?

Confirm whether your vendors use that provider, then rotate any credentials stored with it or pasted into affected AI tools, revoking them at the issuing service. Review the affected content, watch for unusual API usage and billing spikes, and involve legal on notification duties.

When should we use private or hosted inference?

Consider it for workloads involving regulated, confidential or proprietary data, such as legal, financial, health or source code. Running models in an environment you control lets you set retention, logging and endpoint protection yourself instead of inheriting a vendor's subprocessor chain.

Infonaligy helps organizations across the country assess AI vendor and inference supply chain risk, delivered remotely from our Dallas–Fort Worth home base, with on-site support in the cities on our locations page.

Know every link in your AI supply chain

Find out where your AI data really goes.

Book a 30-minute assessment and we'll map the AI features and inference subprocessors your teams use today, review retention, training and breach notice terms, check how secrets and sensitive data reach AI tools, and outline a fourth-party incident playbook.

DFW home base · remote nationwide · governed by default · 800-985-1365