Notion says vendor Fireworks AI security incident exposed Notion AI chat transcripts for fewer than 1% of workspaces

Notion says its AI inference vendor, Fireworks AI, had a security incident that exposed Notion AI chat transcripts (both inputs and outputs) for fewer than 1% of Notion workspaces. Fireworks had contractually promised zero data retention, yet data was retained on its endpoints. As of 3 November Notion has disconnected from Fireworks, deleted the data held there, halted training runs, disabled the Insights feature and stopped serving open-weight models through Fireworks. Affected owners have been notified and can download CSVs of their data. The extraction does not say how the data was exposed, who accessed it, or for how long, and no primary links are listed.

Affected customers should treat the transcripts as exposed and rotate any credentials, API keys or secrets pasted into Notion AI, as Notion advises. The wider lesson for anyone buying AI services is that a contractual zero-data-retention promise is not the same as verified behaviour, so vendor assurances need testing, not just signing. Practical decisions this enables: audit which inference providers and subprocessors sit behind your AI tools, ask for evidence of deletion, and limit what staff paste into chat features. Direct impact is limited by the small share of workspaces hit, but the affected data could be sensitive.

Headlines about a "Notion breach" can overstate it: the exposure sat with a third-party vendor, and fewer than 1% of workspaces were affected, though 1% of a very large user base may still be many organisations. Missing context is the cause of the retention, whether data was actually accessed by outsiders, and Fireworks' own account, so claims of wider industry failure are speculation at this point.

1. 01A Fireworks AI statement or post-incident report explaining why data was retained, how it was exposed, the time window, and whether other Fireworks customers were affected.
2. 02Notion's follow-up on its vendor audit, including any named changes to inference providers, retention verification, or the return of Insights and open-weight models.
3. 03Regulator or customer reaction, such as data protection authority notices, customer contract disputes, or other AI vendors publishing evidence of zero-retention compliance within the coming weeks.