A conversation with an AI can contain the work behind the work: a draft, a customer problem, a fragment of an internal document. Notion and Cursor’s latest accounts of the Fireworks incident show how a stored copy can become a separate exposure. The useful question for a team is where those copies go after the answer arrives.
Based on 7 reviewed sources · See sources
A chat without a name can still identify someone
Notion’s bulletin says fewer than 1% of workspaces were affected. The records included chats used by Insights, its tool for analysing AI usage. Notion disconnected Fireworks on 3 November, deleted its data there on 4 November and contacted affected customers on 9 November. A 13 November update says it reported exposed third-party API keys to their issuers the previous day; exposed Notion keys had been revoked on 11 November. Insights is disabled.
Its incident FAQ makes “unattributable” precise: most records lacked automatically attached identity details, but full user messages and AI replies were included. Names or secrets inside the text could still identify someone. Selected page text and fragments quoted in answers could also appear; whole pages and chat attachments were not included.
Training, generation and storage are different questions
Notion explains that its Enterprise zero-retention commitment covers requests sent to model vendors to generate answers. Features such as history and Insights require stored conversations and fall outside that commitment. Separately, Notion says Fireworks had contractually promised zero retention for its platform use, but the investigation found retained data at the endpoints Notion used.
The current privacy guide also distinguishes plans and models. Most Enterprise generation models offer zero retention; exceptions are off by default and require owner opt-in. Free, Plus and Business provider retention is 30 days or fewer. These are current published rules, not proof of which settings applied to every affected chat.
Read the promises separately
Notion’s current privacy guide describes several different controls.
| Promise or control | What it addresses |
|---|---|
| No training by default | Whether customer data is used to train models. |
| Generation retention | How long a model provider keeps a request and answer; plan and model conditions apply. |
| Workspace permissions | Which content the AI may use when answering. Access permission does not describe retention. |
Cursor traces a different chain from old records
Cursor’s 13 November report describes an exposed vendor key used to copy historical coding-agent records, mostly from 2025, from sessions outside Privacy Mode that were permitted for training. Some contained unredacted credentials. Cursor says fewer than 0.01% of customers had API credentials exposed; stolen credentials then enabled access to selected internal services.
Cursor also attributes the 29 October malicious Claude Code extension copies in its marketplace cache to a stolen credential. Anthropic did not publish them. At most 5,000 users potentially received those specific builds during roughly 11 hours; that is an exposure ceiling, not a confirmed infection count.
Cursor links its incident report and remediation steps. The account above is the company’s investigation, not an independent forensic finding.
View the original post on X ↗The company says Privacy Mode prompts and code, active inference requests and customer code repositories were unaffected. Moving inference away from Fireworks was precautionary. Its original update post points to that investigation.
Follow the affected record
Fireworks’ 9 November bulletin reports impacts to limited datasets and other account information, while saying it had detected no impact to its named Inference, Training SDK and Nexus services. Its investigation remains open. That narrower service statement should not be stretched into an assurance about every stored dataset.
Affected Notion users can review records through the workspace sidebar’s ? menu → Security incident review. Notion recommends rotating exposed credentials. If neither user nor owner was contacted, Notion says the workspace has not been identified as affected; investigation continues.
For Cursor users who installed the affected cached extension versions, its original security notice supplies the specific remediation route and warns that uninstalling alone may be insufficient. This reporting has not inspected customers’ exposed records or independently reproduced the attack.
For a European agency handling a client’s material, the buying question is concrete: which supplier receives the conversation, for what purpose, and what record remains? The same distinction matters when supervising ongoing AI work: a useful answer and a clear account of what happened are separate deliverables.
Sources
- Fireworks AI incident: timeline and 13 November update
Notion · Announcement ·
- Incident review and retention FAQ
Notion Help · Documentation
- Notion AI security and privacy practices
Notion Help · Documentation
- An update on our recent security incident
Cursor · Announcement ·
- Update on October 2026 security incident
Fireworks AI · Announcement ·
- Claude Code extension security notice
Cursor · Documentation
- Cursor’s original incident-update post
Cursor on X · Announcement ·
Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.



