Zero Data Retention Covers a Request Path, Not a Vendor. The Fireworks Breach Reached the Paths Outside It.
Notion had a zero-retention contract and its AI chat transcripts were still in a breached dataset. Three promises with three scopes explain how.
Notion had a contractual zero data retention commitment from Fireworks AI, the inference provider that ran some of its open-weight models. Fireworks documents zero data retention as its default for inference. On 8 November 2026, Fireworks handed Notion a dataset that an intruder had reached, and it held Notion AI chat transcripts from under 1% of Notion workspaces, according to Notion's advisory.
All three statements can be true at once, and that is the lesson for anyone who ships an AI feature on top of a hosted model. Zero data retention describes what happens to one request on one path. It does not describe a vendor, and it does not describe your product.
Most coverage so far addresses Notion customers: check your notice, rotate your keys. This piece is for the team in Notion's position, the SaaS company that calls an inference provider and has told its own customers something reassuring about retention.
What happened, in dates
Fireworks has published two bulletins. The 9 November update describes two separate events: misuse of internal credentials found on 6 October, and a second intrusion on 24 October that began at a third-party cloud provider hosting Fireworks research machines. The Notion and Uber disclosures followed from those.
| Date (2026) | Event | Source |
|---|---|---|
| 6 Oct | Fireworks finds unauthorised use of internal credentials; environment variables in evaluation jobs exposed for some accounts | Fireworks |
| 9 Oct | First Fireworks bulletin. Uber is told its account with a provider was accessed | Fireworks; Uber notice |
| 13 Oct | Uber determines driver personal data was accessed | Uber notice |
| 23 Oct | Third-party cloud provider hosting Fireworks R&D machines has a security incident | Fireworks |
| 24 Oct | Intruder reaches Fireworks development machines and runs an unauthorised workload. Detected the same day | Fireworks |
| 2 Nov | Fireworks alerts Notion to an exposed dataset that first appears to hold only synthetic test data | Notion |
| 3 to 4 Nov | Notion cuts traffic to Fireworks and deletes its data there | Notion |
| 6 Nov | Fireworks tells Notion that customer data may be involved | Notion |
| 8 Nov | Fireworks confirms the dataset; Notion finds AI chat transcripts used for Insights | Notion |
| 9 Nov | Fireworks publishes its update; Notion notifies customers | Both |
| 10 Nov | Date on Uber's breach notice to the California attorney general | Press reports |
Fireworks says the second intrusion reached resource metadata, billing information and, for certain customers, a subset of datasets and incomplete model weights. It says it detected no impact to its Inference, Training SDK or Nexus products, which it describes as zero data retention by default. Uber confirmed to Gizmodo that Fireworks was the provider behind its own notices, which cover driver's licence information for at least 7,870 people across three US states. Uber has not said why that data was with a model host.
Three zero data retention promises, three scopes
Read the documents side by side and there are three separate promises in this stack, each with a different boundary.
| Layer | What it says | What it leaves out |
|---|---|---|
| Provider default | Prompts and generations for open-model inference live in memory for the length of the request | Batch jobs, stored responses, training datasets. These store content by design |
| Provider enforced policy | An account-wide setting that rejects any request or job that would persist customer content | Enterprise accounts only, off until an admin enables it, and it does not delete content stored earlier |
| Application promise to customers | Requests sent to model vendors to generate AI responses are not retained by those vendors | Features that need stored requests and responses, such as chat history and usage analytics |
The first row comes from Fireworks' data handling documentation, which lists the exceptions plainly: the Batch API needs dataset storage, and the Responses API stores conversations for up to 30 days unless the caller sets store to false. The second row is from the same documentation, with the defaults as VentureBeat reported them. The third is how Notion describes its own Enterprise offering, as summarised by ComplexDiscovery.
Notion's Insights feature gave workspace admins a view of how their teams used Notion AI. It worked by sending stored chat transcripts to Fireworks for classification. So the feature sat outside the third promise by Notion's own definition, and classification over a stored dataset is the kind of job that sits outside the first.
The features that store are not the ones you call inference
Ask an engineer where the product sends prompts and you will hear about the chat completions call. That path is usually the clean one. Storage enters through the features built around it, and most of them were added later by a different team for a different reason.
A working list, drawn from how hosted model platforms are documented:
- Batch inference. You upload a file of requests and collect a file of results. Both files are datasets at rest on the provider.
- Stored responses and conversation state. Some APIs keep the conversation server-side by default so the next call can refer to it.
- Evaluation jobs. Test sets are often sampled from production traffic, and the job may carry your API keys as environment variables. That was the first Fireworks incident.
- Fine-tuning and distillation. Training data is uploaded and kept for the life of the job, often longer.
- Analytics over conversations. Usage dashboards, topic clustering, quality scoring. This was Insights.
- Debug logging and tracing. A flag turned on during an incident and never turned off.
Each of these takes content that was transient on the request path and writes it to disk somewhere. Each is also easy to describe internally as "not part of inference", which is how it escapes a retention review that was scoped to inference.
A promise drifts. A policy that rejects requests does not.
A contractual commitment tells you who is at fault afterwards. It does nothing at the moment a new feature starts writing customer content to a provider's storage. Nobody at either company has to make a decision for that to happen; a batch job only has to be submitted and accepted.
An enforced policy changes the failure. With the Fireworks account-level setting enabled, chat completions go through and batch jobs, evaluation jobs and dataset uploads are rejected, per VentureBeat's reading of the documentation. The engineer who tries to ship a storage-backed feature gets an error on the first call. That error forces the conversation about retention to happen before the data moves.
The practical pattern is two accounts or projects per provider. One has the retention policy enforced and carries everything you have promised customers will not be stored. The other is for work that needs storage, and anything that runs there needs a named owner, a retention period and a line in your customer documentation. If a provider offers no enforceable setting, the split still helps, because the storage-capable credentials are held by fewer services.
Two cautions. Turning a policy on does not remove what is already stored, so audit existing datasets and stored responses separately. And an absent setting usually means off. Read the configuration; do not infer it from the marketing page.
The notice chain took 15 days, and the first answer was wrong
Fireworks detected the second intrusion on 24 October. Notion heard about a possibly exposed dataset on 2 November, nine days later, and was told it looked like synthetic test data. Confirmation that real transcripts were involved came on 8 November, 15 days after detection. Notion then notified customers within about a day.
That last interval is the one Notion controlled, and it was short. The other 15 days were upstream. Uber's sequence has the same shape: it determined on 13 October that driver data was accessed, and the California notice is dated 10 November.
Under GDPR, a processor must tell the controller about a breach without undue delay, and the controller has 72 hours from becoming aware to tell the regulator. India's DPDP Rules set a 72-hour window for the detailed report to the Data Protection Board once those obligations take effect. None of these clocks say much about a subprocessor two layers down that needs a fortnight of forensics to know whose data was in a bucket.
The Cloud Security Alliance research note on this incident recommends flowing breach-notification timelines down to the inference layer in contracts. Add two operational points. First, decide in advance what you will do on a "probably nothing" alert from a provider. Notion cut traffic the morning after an alert that described synthetic data, which looks right in hindsight. Second, agree who preserves what. Deleting your data from a breached provider is a sound reflex, but the provider's copy is also the evidence of what the intruder could reach.
Transcripts are credential stores
After reviewing the dataset, Notion revoked Notion API keys it found in the affected messages and reported third-party keys to the vendors that issued them. People paste secrets into assistants: a failing curl command, a connection string, a config file they want explained. Any system that stores chat transcripts stores some of those.
So the classification for conversation logs should be the one you use for a secrets store, with the access controls and retention limits that implies. "Unattributable" helps less than it sounds. A transcript with no workspace label can still name a client, a colleague or a project in its own text.
Five questions to answer about your own product this week
- For each AI feature, which provider endpoints does it call, and does any of them write the request or the response to storage? List features, not vendors.
- Does the provider have a setting that rejects storing operations, and is it on for the account your production keys belong to? Check the setting itself.
- Does your customer-facing retention statement name the features it excludes? If analytics or history sit outside it, say so where a buyer will read it.
- What does your provider contract say about how fast they must notify you, and what happens on a provisional alert?
- What is already stored? Datasets, stored responses, old evaluation sets and secrets attached to jobs all outlive the feature that created them. Rotate keys at the issuer, since deleting a stored copy does not invalidate the key.
Fireworks says its investigation is continuing and its findings may change, and BleepingComputer reports that Notion is now auditing its critical vendors for retention compliance. Expect enterprise security questionnaires in 2027 to ask for retention by feature and by processing path. The teams that can answer from a configuration export will close those reviews faster than the ones answering from a contract clause.
Frequently asked questions
Related reading
Developer Trust in AI Is Conditional: Only 6.6% Trust It With Important Decisions
Stack Overflow's 2026 survey: 80% use AI for an hour a day, 48% trust it only when they can verify it, 6.6% trust it with important decisions. Verification cost sets the ceiling.
OpenAI Reports 948 Linux Privilege-Escalation Findings. The Number Is a Review Queue, Not Your Exposure.
OpenAI says an unreleased cyber model produced 948 Linux kernel privilege-escalation results for about $1.2M. Here is what that number measures, and what to check on your own fleet.
Idempotency Keys for Agent Tool Calls Need a Third State: "We Don't Know"
Most idempotency advice for LLM agents stops at "hash the arguments and keep a ledger". The failure that costs money is the crash after the side effect and before the record.