The takeaway
Zero data retention is a per-feature setting, not a vendor property. If an agent classifies or evaluates conversations, the transcripts sit at rest on a third party's infrastructure.
Why it matters for builders
Audit inference vendors feature by feature. Chat completions can be ephemeral while batch jobs, evaluation runs and analytics pipelines persist the same content. Assume pasted credentials are compromised after any vendor incident and rotate them.
Notion AI Chat Transcripts Leaked Through a Vendor Breach
Fireworks AI has confirmed that an attacker reached a development system through a third-party cloud provider and left with customer billing information, secrets and datasets. One of those customers was Notion, which now says the stolen files included Notion AI chat transcripts belonging to fewer than 1% of its workspaces.
What happened
On November 2, Fireworks told Notion a small set of data had been accessed. Notion checked and found only synthetic testing data. By the morning of November 3 it had cut every connection to the vendor, and on November 4 it deleted all of its data from Fireworks systems. On November 6, Fireworks warned that customer data may have been accessed. On November 8 it confirmed the final affected dataset, and on November 9 Notion began contacting impacted customers.
That dataset held chat transcripts from the Insights feature, which analysed how teams use Notion AI. The company disabled Insights and stopped serving open-weight models through Fireworks entirely.
Much of the data is unattributable. Notion says most transcripts carried no label tying a conversation to a user or workspace, and no tool call results. Notion told customers to treat the data as exposed and rotate any password or API key they had ever pasted into Notion AI. On November 11 it revoked Notion API keys found in the affected messages, and on November 12 it reported exposed third-party API keys to the vendors that issued them.
Fireworks attributes the activity to financially motivated criminals reselling unauthorised model access to illicit routing gateways, a pattern Google's threat intelligence team calls LLMJacking. The company says a separate October incident, traced to a third-party cloud provider hosting its R&D machines, produced the broader data impact.

Why it matters
The uncomfortable part is not the intrusion. It is where the data was sitting.
Fireworks advertises zero data retention by default. Its own documentation narrows that: prompts and generations live in volatile memory only for the length of a request, but stored responses, batch jobs, evaluations and training all persist content. Notion's transcripts existed because a classification job had to keep them. The retention default was never the guarantee people assumed it was.
For teams routing production traffic through an inference vendor: zero retention is a per-feature setting, not a property of the company. If an agent classifies, evaluates or logs conversations, the content is at rest somewhere, inside someone else's perimeter.
What builders should do
- Inventory vendor features, not endpoints. Batch inference, evaluation and storage-backed analytics are where content survives.
- Treat transcripts as secret-bearing. Any credential that ever touched a model chat should be rotated after an incident.
- Shrink the blast radius early. Notion severed connectivity a day after the first alert, before customer impact was confirmed.
Notion says affected workspaces have been contacted, and its investigation is ongoing.
The Automation Brief
Read 5 AI stories instead of 50.
The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.
No noise. Unsubscribe anytime.
Editorial notes
Stefan Trbojevic
n8n Lab Editorial
13 November 2026
13 November 2026
Sources
AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.




