Notion AI Data Breach: Check Your Chats, API Keys and Exposure (November 2026)

当ページのリンクには広告が含まれています。
Notion AIの会話記録の一部が委託先の不正アクセスにより流出した可能性を説明する概念図
ノートンストア

Did the Fireworks AI breach expose your Notion AI chats or API keys? Notion says a security incident at its third-party AI provider affected fewer than 1% of Notion workspaces. The affected dataset included some Notion AI chat transcripts used for its Insights feature. Notion has contacted identified affected customers. If neither you nor your workspace owner has received a notice, Notion says no incident-specific action is necessary at this time, although its investigation continues. Last checked: November 15, 2026.

Start here: Verify your workspace notification, review the affected transcript through Notion’s official Security incident review interface, and revoke or rotate any credentials you find in the affected messages. Notion reported revoking exposed Notion API keys on November 11 and notifying third-party key issuers on November 12. This does not mean your other providers’ credentials have all been revoked for you.

広告・PR
目次

What happened in the Notion AI / Fireworks AI security incident?

The compromise was at Fireworks AI, a vendor used by Notion, rather than a confirmed intrusion into Notion’s core systems. Fireworks reported unauthorized activity involving its environment, including additional customer-data impact from an October 24 incident. Notion previously used Fireworks for data-classification jobs powering the Insights feature and to serve certain open-weight models.

QuestionVerified position as of November 15, 2026
Who may be affected?Fewer than 1% of Notion workspaces; a global count of affected people or transcripts was not disclosed.
What data?Selected Notion AI chat transcripts used for Insights, including user prompts and assistant responses.
Were Notion pages or attachments broadly exposed?Notion says the dataset did not automatically include whole pages, spreadsheets or image attachments. However, selected page text or excerpts quoted in the AI response may be included.
Was Notion itself breached?Notion says the incident was at its vendor, and no other Notion features were affected.
What should you do?Check notification and the incident-review data, then rotate exposed credentials and check related service logs.
Is this resolved?Notion and Fireworks continued investigations and monitoring as of the cited updates.

Most affected transcripts lacked metadata automatically identifying the writer or workspace. But the text itself can reveal an identity or organization—for example, an employee may have pasted a customer name, email address, internal procedure or secret. The absence of automatic identifiers does not make the content harmless. Source: Notion security advisory.

How to check whether your Notion workspace is affected

Notion has published a specific review workflow for impacted workspace owners and members. Do not enter your credentials on an unverified incident-notice website: open Notion directly from your usual bookmark or app first.

  • Step 1 — Check the incident notice: Ask the workspace owner whether Notion sent an affected-workspace notification. Notion says affected customers were contacted directly; absence of a notice currently means it has not identified your workspace as affected.
  • Step 2 — Open the review: In an affected Notion workspace, use the sidebar ? → Security incident review. Download the files made available for your account. The option is for affected owners or users, not necessarily every Notion account.
  • Step 3 — Inspect the transcript: Open affected_content.csv in a controlled environment. The text field holds message content, while message_role distinguishes user messages and AI responses.
  • Step 4 — Inventory sensitive content: Look for passwords, API tokens, private keys, internal links, customer names, email addresses, contact details, confidential snippets and anything that identifies a client or organization.
  • Step 5 — Take action: Revoke or rotate affected credentials at the issuing service, update applications or integrations using them, inspect relevant audit logs, and follow your organization’s incident-response process.

How to analyze the annotated CSV efficiently

For certain Business and Enterprise workspaces, Notion also provides affected_content_annotated.csv. Filter __analysis_secret_categories for values other than [] to find potential credentials; __analysis_pii_categories flags categories such as person, email, phone, address and account number. These flags are screening aids, not proof of everything that was or was not sensitive. Review the original message text manually, especially for your organization’s own terminology.

Notion explains that fields beginning __metadata or __analysis were added for post-incident review and were not part of the exposed dataset. The CSV files themselves can contain sensitive information. Restrict access to the downloaded files, and remove local review copies when your organization no longer needs them. Source: Notion incident FAQ and data-review instructions.

Were passwords, Notion API keys and other credentials exposed?

Potentially, if they appeared in an affected chat. Notion specifically warns that messages could contain a user-supplied password, API token or other secret. It says it revoked exposed Notion API keys on November 11. On November 12, Notion notified the vendors that issued third-party API keys found in affected user messages, and published that development on November 13.

You should not assume that the notification to an outside vendor automatically rotated your token. The owner of the token must check it with its issuer and revoke or rotate it as needed. If a Notion integration stops authenticating, review whether its Notion API key was among those revoked and issue a replacement through the authorized integration owner.

Priority order for IT and security administrators

  • Critical credentials first: Production cloud accounts, administrator tokens, payment-service keys, SaaS integrations, source-code access and customer-data credentials.
  • Invalidate the old secret: Simply deleting it from a chat or note does not revoke a usable credential. Rotate it at the service that issued it.
  • Update dependencies: Replace the secret in authorized secret managers, CI/CD jobs and integration settings; verify that jobs and connectors work after the change.
  • Check for abuse: Review identity, cloud and application audit logs for suspicious access, unexpected API usage or data downloads. Preserve evidence when an incident is suspected.
  • Coordinate notifications: Assess personal-data or contractual notification obligations based on what was actually present in the affected records, not solely on a headline.

What information was—and was not—in the affected chats?

Notion’s detailed FAQ distinguishes four parts of an AI conversation: automatically supplied metadata, user chat text, tool usage, and the assistant’s output. The affected transcripts included the full user message and full assistant response. They could also contain selected text from a page or content the assistant quoted in its answer.

Notion says the affected data did not automatically include the rest of a page or attachments such as image files and spreadsheets. Most affected transcripts did not include the results of the assistant’s tool calls, except for content that the final response quoted. AI meeting-note data was not affected as standalone meeting-note content unless it was entered into or referenced in the affected chat. These distinctions help scope the review, but do not justify skipping a manual check.

Timeline: October attack, November customer notifications

Date in 2026Verified event
October 23–24Fireworks reported an incident at a third-party cloud provider hosting its R&D machines; unauthorized activity on October 24 led to additional customer-data impact.
November 2Fireworks initially notified Notion of a limited accessed dataset. The initial assessment described synthetic test data rather than customer data.
November 3–4Notion stopped traffic to Fireworks by November 3 and removed its data from the vendor’s systems on November 4.
November 6–8Fireworks raised the possibility of customer-data access on November 6 and confirmed the final impacted dataset on November 8.
November 9Notion began contacting identified affected customers.
November 11Notion reported revoking Notion API keys found exposed in the affected messages.
November 12–13Notion published the security advisory and reported contacting issuers of exposed third-party keys on November 12; the related update appeared November 13.

Fireworks separately disclosed evidence pointing to financially motivated LLMJacking, which involves illicit use or resale of access to language-model APIs. That broader finding should not be treated as proof of exactly how any given Notion AI transcript was retrieved or misused. The investigation is ongoing. See Fireworks’ November 9 security bulletin.

Examples of downstream impact disclosed by Japanese organizations

The incident also matters to people who never personally used Notion AI: employees at another organization may have included their contact information in a work-related prompt. The following figures are company-specific potential exposure counts, not global breach totals or necessarily unique individuals.

OrganizationPublic disclosure
ONE CAREER (Nov. 12)Reported up to approximately 2,300 personal-information entries in affected Notion AI chats. Its own employment-service systems were not breached in this event.
For Startups (Nov. 13)Reported up to approximately 1,900 potentially affected entries and planned to notify identified individuals.
Keio University (Nov. 13)Confirmed it had been notified about possible impact on its use of Notion, with assessment continuing.
JMAM (Nov. 13)Said affected chat records could contain personal or business information concerning 51 companies in the Tokyo area. This is a count of companies referred to, not 51 confirmed breaches.
Sankei Shimbun (Nov. 14)A business partner’s use of Notion raised a possible exposure of system-related information. Sankei iD was temporarily suspended on Nov. 13 and resumed on the morning of Nov. 14 following safety checks; investigation continued.

If your company is named in a disclosure, read its own notice and await direct communication where promised. Avoid phishing messages claiming you must immediately enter your password or payment details into a new incident portal.

Did someone also claim to sell a huge Notion database?

That is a separate claim from the Fireworks AI incident. On November 13, Notion published an advisory addressing an allegation that a large Notion dataset was being offered for sale. Notion said its investigation found the shared sample inconsistent with genuine Notion system data and no evidence that the allegation reflected a breach of Notion customer data. Do not merge that unsubstantiated claim with the documented vendor incident. Source: Notion statement.

FAQ: questions users and administrators are asking

I never received an email from Notion. Do I have to reset every password?

Notion says no incident-specific action is required at present if neither you nor the workspace owner has been contacted. This does not rule out unrelated security issues or future updates as the inquiry continues.

Could someone identify me if the transcripts had no user metadata?

Yes. A name, business detail, email address or other identifying content in a prompt or assistant response may still identify a person or company. That is why Notion asks affected users to inspect the content itself.

Was my entire Notion workspace or every database downloaded?

Notion has not described an exposure of all workspace pages or databases through this incident. The identified dataset involved selected chat transcripts for Insights; quoted snippets and selected page text could appear inside those transcripts.

Does turning off AI training prevent vendor incidents?

No. A training-use setting governs whether data is used for model improvement; it does not by itself prevent processing by infrastructure vendors, operational analysis or failures in access controls. Restrict what is submitted to AI systems, manage secrets outside chats and review vendor data flows and retention commitments.

Can I delete the exposed information from the attacker’s copy?

Deleting a message now does not reliably revoke access to a copy already obtained or exposed. Prioritize invalidating any secrets and following the incident-review and notification process.

Bottom line and official sources

If affected, review your transcript and rotate secrets first. If you have no notification, Notion currently does not require special action. The precise global number of exposed users and any complete downstream misuse assessment remain unconfirmed as of November 15, 2026. Avoid confusing possible exposure with proven downstream identity theft.

Editorial note: This article summarizes publicly documented facts and general remediation guidance. The precise affected-user scope and individual organization findings may change as investigations progress. Published November 15, 2026; source review completed the same day (JST).

よかったらシェアしてね!
  • URLをコピーしました!
  • URLをコピーしました!

この記事を書いた人

はじめまして、「ぴろりん」です。
このブログでは、IT運用管理・PC修理・情報セキュリティ対応・資格学習など、実務で直面した課題とその解決策を発信しています。
同じような悩みを抱える方のお役に立てれば幸いです。

目次