AI Governance
Zero data retention by default does not tell you where your AI vendor's vendor stores data
Notion AI chat transcripts were in a dataset taken from Fireworks AI, an inference provider whose core inference is zero data retention by default. The retention question that matters is feature by feature, and it reaches one layer further than most vendor reviews go.
November 17, 2026 · 6 min read

Notion AI chat transcripts were in a dataset that attackers took from Fireworks AI, one of Notion's AI infrastructure vendors. Notion's security advisory says the transcripts were the ones used for its Insights feature, and that fewer than 1% of Notion workspaces were affected. Fireworks, for its part, says its inference product is "zero data retention (ZDR) by default."
Both statements can be true at once, and that is the point of this post. If you sell software with AI features to health systems, your buyers are going to ask where their data goes. "Our model provider has zero data retention" is not a complete answer, and after this month, a careful buyer will know it.
What Notion and Fireworks have said
Notion's advisory lays out the timeline. On November 2, Fireworks told Notion a small set of data had been accessed, which turned out to be synthetic testing data. By the morning of November 3, Notion had disconnected from Fireworks. On November 6, Fireworks said customer data might be involved, and on November 8 it confirmed the dataset. Notion found it included Notion AI chat transcripts used for Insights, which for most affected workspaces were "unattributable," meaning not labeled with the user or workspace. Notion says it used Fireworks for classification jobs behind Insights and as a general inference provider for open-weight models. It has disabled Insights and stopped serving models through Fireworks.
Notion told affected customers to treat the data as exposed and to rotate any passwords or keys they had pasted into Notion AI. On November 11, it revoked exposed Notion API keys found in affected messages, and on November 12 it reported third-party API keys it found to the vendors that issued them.
Fireworks' November 9 update says a third-party cloud provider hosting Fireworks R&D machines had its own incident on October 23, and that the next day an unauthorized party used it to reach Fireworks development machines and run a workload that led to additional customer data impact. Fireworks lists the affected categories as resource metadata, billing information, and, for certain customers, a subset of incomplete model weights and datasets. It says it has not detected impact to Inference, Training SDK, or Nexus, and it disabled self-serve dataset uploads and downloads used by Batch Inference while it investigates.
Where the gap is
VentureBeat's analysis does the useful work of reading Fireworks' own documentation next to the disclosures. It quotes Fireworks' policy page saying that by default prompt and generation data "exist only in volatile memory for the duration of the request," and that "a few features do store customer content, such as stored responses, batch jobs, and training." It also reports that the separate account-level Zero Data Retention policy, which rejects jobs that would store content, is available on Enterprise accounts and is off by default.
VentureBeat is also careful to say what nobody has disclosed: neither company has said which Fireworks feature stored Notion's transcripts, and nothing shows the policy would have stopped the intrusion. I am going to respect that limit. I do not know why the transcripts were stored, and I am not going to guess.
What the public record does show is the shape of the problem. A vendor can truthfully say its inference path keeps nothing, while a different feature on the same platform, used for a different job, stores content at rest. A classification job is a reasonable thing to send through a batch process. A buyer asking "do you retain our prompts?" may get an honest "no" about the chat path and never hear about the job.
The question to ask your AI subprocessors
If you are a HealthTech company, the people asking you these questions are hospital security and privacy teams. The people you should be asking are your own AI subprocessors. Put the answers on one page, because that page is what you will hand to a buyer.
For each AI vendor that touches customer data, ask in writing:
- Which of our features call you, and which of your products or endpoints does each one use? List chat completions, batch jobs, fine-tuning, evaluation, embeddings, and anything else separately.
- For each one, is any customer content written to disk, a queue, a log, or object storage, even temporarily? For how long?
- Is your zero data retention a default, a contract term, or an account setting? If it is a setting, show us that it is turned on for our account.
- Which of your own subprocessors, including cloud hosts, can reach stored content or the systems that process it?
- If stored content is accessed without authorization, how soon will you tell us, and will you tell us which of our features the data came from?
- When we stop using a feature or end the contract, how do we confirm deletion?
Then turn the answers into an attestation your buyers can rely on: one row per AI feature, naming the vendor, the endpoint type, whether content is stored, for how long, and where.
Secrets in prompts are a separate problem
The detail in Notion's advisory that I would put in front of every product team is the API keys. Users pasted credentials into an AI chat, those chats were stored, and the keys had to be revoked and reported. That will happen in your product too. People paste whatever is on their screen.
Two controls help. Scan prompts for obvious secrets before they leave your system, and redact or block them. And treat any feature that stores prompts as a store of secrets for retention and access purposes, because some of the time it will be.
What to do before a buyer asks
Pull your list of AI vendors and the features that call each one. If you cannot say, for every feature, whether content is stored at the vendor and for how long, start there. The Fireworks disclosures show the answer can be different for two features that look identical from the outside. A health system's vendor review will get to that question eventually, and it is better if your one-page attestation is already written when it does.