Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentData exposure

Notion: the Fireworks AI breach exposed Notion AI chat transcripts from fewer than 1 percent of workspaces

MediumNotionFireworks AI

Notion published an advisory on November 10, 2026, about a security incident at Fireworks AI, a third-party provider it used for inference on open-weight models. Fireworks told Notion on November 2 that a small set of data had been accessed, which Notion found was synthetic test data; on November 6 Fireworks said attackers may have reached customer data, and on November 8 it confirmed the affected dataset. The data was Notion AI chat transcripts used for the Insights feature, from fewer than 1 percent of Notion workspaces; for most affected workspaces they carry no metadata that ties them to a workspace, user or company. Notion says this was not a breach of its systems and no other feature was affected. Fireworks had given a contractual commitment to zero data retention, but data was retained from the endpoints Notion used. Fireworks says a third-party cloud provider hosting its research machines had an incident on October 23, and on October 24 an unauthorized party reached its development machines and ran an unauthorized workload; it links the activity to a financially motivated group that sells LLM access through illicit routing gateways. Neither company gives a number of affected customers. On November 13 Cursor said an access key it had issued to Fireworks, exposed by Fireworks, let an attacker copy Cursor coding agent traces from a shared storage bucket, which led to the tampered Claude Code extension in Cursor's marketplace.

Replay the attack step by step in the Learning Center

How it happened

  1. Attackers reached Fireworks AI's development machines through a third-party cloud provider and ran their own workload.
  2. Data from the inference endpoints Notion used had been retained there despite a zero retention commitment.
  3. Notion AI chat transcripts for the Insights feature were in that data.

Impact

Notion AI chat transcripts from fewer than 1 percent of workspaces were exposed, including any secrets users had pasted into them.

The fix

Notion disconnected from Fireworks, deleted its data there, disabled Insights, stopped serving open-weight models through Fireworks, contacted affected customers on November 9 and on November 11 revoked Notion API keys found in their messages. It asks customers to treat the data as exposed and rotate any credential pasted into Notion AI; workspace owners can review the affected content in the desktop app. Fireworks revoked and replaced credentials, isolated its network, tightened storage access and added logging.

Lessons

  • An AI feature's model provider holds your prompts too; ask where they are kept and check it.
  • A zero retention clause is a promise, not a control you can see working.
  • Keep secrets out of AI chats; if one went in, rotate it.

Sources