Zero Data Retention Was True. The Breach Hit the Three Exceptions.

On 9 November, Notion published a security advisory. A breach at its AI vendor, Fireworks AI, had exposed Notion AI chat transcripts from fewer than 1% of Notion workspaces. Notion's own systems were not breached. It disconnected from Fireworks on 3 November, deleted the data held there, and stopped sending traffic.
What lifts this above a news brief is the retention policy. Fireworks' documentation states, in effect, that prompt and generation data exist only in volatile memory for the duration of the request. That is zero data retention, and it was true of the endpoints most customers think about.
The same documentation adds that a few features do store customer content: stored responses, batch jobs, and training. The breach reached the stored side.
The timeline, as disclosed
Two separate intrusions, and the second one is the one that mattered.
On 6 October, Fireworks detected unauthorised use of internal credentials. The company later said environment variables, including API keys and tokens stored as secrets, were captured in certain affected accounts.
Then a third party got involved. A cloud provider hosting Fireworks research and development machines had a security incident on 23 October. On 24 October, an unauthorised party used that access to reach Fireworks development machines and run an unauthorised workload, which led to additional customer data impact.
The material taken across both: resource metadata, billing and payment records, a subset of some customers' datasets, and incomplete model weights. Fireworks says it does not store customer card numbers, and that Inference, Training SDK and Nexus are unaffected because they are zero data retention by default. The CEO's wording adds a parenthesis the security update does not: Inference, excluding batch inference.
Fireworks describes the actor as financially motivated, aiming to sell access to large language models through illicit routing gateways. That practice has a name, LLMjacking, and Palo Alto Networks Unit 42 was brought in to help.
Notion's four days
The Notion side is a case study in how a vendor incident propagates.
On the evening of 2 November, Fireworks told Notion that a small dataset had been accessed. That first dataset contained only synthetic test data.

By the morning of 3 November, Notion had disconnected from Fireworks, deleted the data it held there, and stopped sending it traffic. That is a fast response, and it did not prevent the later exposure, because the data was already stored at Fireworks before the connection was cut.
On 6 November, Fireworks said attackers may also have reached customer data. On 8 November it confirmed. The dataset included Notion AI chat transcripts used for the Insights feature, which shows customers how their teams use Notion AI. Notion published on 9 November, began contacting affected customers, and on 11 November revoked Notion API keys found in affected messages.
Notion also used Fireworks to run open-weight models. That stopped. Insights is off while the investigation continues.
The scope is narrower than the headlines suggested. Fewer than 1% of workspaces. For the vast majority of affected workspaces, the transcripts carried no label identifying the author or the workspace, and they did not include tool-call results.
Narrow is not safe. People paste passwords, API keys and customer details into AI chats, and that text is sensitive whether or not it arrives with a name attached. Notion's own advice is to treat the data as exposed and rotate anything pasted into Notion AI during the window.
Why the retention claim and the breach can both be true
This is the part security teams should take away, and it is a question of scope rather than honesty.
Zero data retention is a statement about a request path. A prompt comes in, tokens come out, and the content lives in memory for the duration. Nothing is written to disk. That is a real property and a genuinely good default.
Storage is a different product surface. Batch inference exists because some jobs are too big to answer in one request, which means inputs and outputs have to be written somewhere until the job finishes. Stored responses exist because customers asked to retrieve them later. Training exists because that is what training is.
So a vendor can hold a true ZDR claim and still hold your data, because the data lives in the services that are not requests. The breach disclosures read exactly that way: the impact was on storage systems, and the batch, dataset and response features are the ones that write to storage.
There is a second wrinkle. The ZDR policy is available on Enterprise accounts, and its default setting reads as off by default. So for many customers, ZDR is a thing you can turn on rather than a thing that is on.
None of that makes Fireworks unusual. Every inference vendor has this shape, because batch and training are where the money is, and both require persistence. The lesson is about the questionnaire. Asking "do you retain my data" gets a yes or a no about a request. Asking "which endpoints persist content, for how long, in which region, and who can read that storage" gets an answer you can act on.
The economics behind it
The motive explains why this keeps happening.
LLMjacking is arbitrage rather than espionage. Stolen inference access gets resold to routing gateways that want to serve model traffic without paying for it. The stolen asset is your token budget and your rate limits, and the damage is a bill plus whatever your data was worth.
The surrounding numbers are ugly. Railway disclosed in late October that attackers decrypted environment variables for up to 100,000 workspaces and ran up about $740,000 in fraudulent AI usage. Wiz counted at least nine AI developer platforms disclosing incidents in October, and traced two stolen enterprise keys to a single vendor, both of which showed identical first unauthorised activity within 30 minutes of the reported credential access.
Thirty minutes. The gap between a credential appearing in someone's logs and someone spending it is now shorter than a lunch break, which means detection windows measured in days are decorative.
What to do differently
Four things, in the order I would do them.
Inventory the stored surface. List every endpoint your product calls where data persists: batch jobs, stored responses, datasets, fine-tuning, evaluation jobs. If you cannot list them, you cannot assess the breach you are about to read about.
Stop treating chat transcripts as low-sensitivity. They are the highest-density secret channel in most organisations, because a human pastes a config file into a chat box to ask a question about it. Whatever retention story applies, the transcript is the crown jewel.
Rotate on disclosure, not on confirmation. Notion deleted and disconnected on 3 November, before anyone knew whether real data was involved, and revoked API keys on 11 November when the scan found them. Both were correct.
Assume the vendor's vendor. The intrusion started at a cloud provider hosting development machines. Your AI vendor's security posture includes companies you have never heard of, and the only way to price that in is to require disclosure of subcontractors and to prefer vendors who will name them.
The uncomfortable summary
Notion did most things right, moved in under twelve hours, and still had chat transcripts exposed. Fireworks published a retention policy that was accurate and a breach that landed in the part it excluded. Both companies told the truth and the outcome was still bad for a small number of customers.
That is the shape of supply-chain risk in AI infrastructure, and it is the reason the story is about scope rather than about dishonesty. A policy can be true and insufficient at the same time. The work is in asking which nouns the policy applies to before you find out from an advisory.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.