Notion AI Chat Data Leak Feared: Three Domestic Companies Report Incident
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever worried, "What if something I told Notion AI was stolen without my knowledge?" In November 2026, that very concern became a reality. This article breaks down what happened, who is affected, and what you should do starting today.
US-based Notion Labs, which operates the cloud workspace "Notion," has announced that Notion AI chat logs may have been leaked. According to reporting by ITmedia AI+, the announcement was made on November 9 local time.
The cause was not Notion itself, but rather unauthorized access to its subcontractor, US-based Fireworks.ai (a cloud provider that runs AI models). Notion has stated that its own systems were not compromised.
The chat logs at risk of exposure are those sent for use by the "Insights" feature, which analyzes Notion AI usage. The content users type directly into the AI is what falls within scope.
The number of affected workspaces is said to be less than 1% of the total. However, the number of chat records involved and the period covered have not yet been disclosed.
The full picture of this incident emerged gradually over several weeks. Here is a summary of the key events:
Minor discrepancies in specific dates exist between the Notebookcheck article and an explanatory article summarizing Notion's announcement. Please refer to Notion's official announcement for accurate dates.
Fireworks has stated that behind the attack was a financially motivated group involved in LLMjacking — a crime in which stolen credentials are used to access AI services without authorization. According to VentureBeat's reporting, this group reportedly sells illicitly obtained AI access to unofficial relay services.
In the initial breach in October, secret information such as API keys (credentials used to access AI services) is said to have been viewed. The subsequent breach resulted in chat logs being exfiltrated as well.
Fireworks has stated that for ordinary inference processing, ZDR (Zero Data Retention: a system that does not store interactions) is the default. So why were chat logs stolen?
VentureBeat points out that there are features that do store content, such as batch processing and dataset uploads. Notion's chats may have been held in such a storage location. However, neither company has clarified which feature was responsible for storing the data.
In Japan, companies using Notion have also issued reports one after another. According to ITmedia, the three companies that reported are as follows:
All three companies are still confirming the details of what was leaked. They have stated they will notify customers individually if their information was included, and are also urging caution regarding suspicious emails or phone calls impersonating their companies or Notion.
One particularly concerning point is that "meeting notes" may also be included. Many users likely asked AI to summarize meeting minutes, and if customer names or discussion details were included, that information may have been exposed externally as-is.
Incidents involving AI are not unique to Notion. According to VentureBeat, security firm Wiz counted at least nine AI development platforms that disclosed incidents in October alone.
In this case, Notion had delegated the processing of AI model operations to Fireworks. A subcontractor invisible to end users had become a conduit for sensitive information.
Here are some key points to verify when using AI services:
The Cloud Security Alliance also recommends asking vendors in questionnaires who can decrypt customers' confidential information.
Consider a sales representative at a small business. They paste meeting notes into Notion AI and ask it to "draft a follow-up email." The notes contain the name of a client contact and discount terms.
If that conversation is leaked, it could damage the trust relationship with the client. And the person themselves has no way of knowing whether they were affected until they receive a notification from their company.
The same applies to engineers. When troubleshooting an error, it's not uncommon to paste an API key along with logs into an AI tool. Notion has stated that it invalidated API keys found within the leaked messages.
So what should you do now? Based on Notion's guidance, here is a summary:
For companies, this is an opportunity to revisit the rules around what information is acceptable to share with AI. The safest approach is to adopt a practice of simply not entering personal information or customer names in the first place.
Notion is directly contacting workspaces that have been affected. If you have not received a notification, no additional action is required at this time. The impact is limited to less than 1% of all workspaces.
No. Notion has stated that its own systems were not compromised, and that the data was held on the Fireworks side.
According to Notion, many of the affected workspaces did not have authorship or workspace information attached to the chats. However, if names or company names were written in the body of the chat, it may be possible to identify the owner from that.
The Insights analytics feature has been suspended while the investigation is ongoing. No timeline for resumption has been provided. Notion has also stopped offering open-weight models (AI models released for anyone to use) via Fireworks.
Companies should define the scope of information acceptable for AI input and communicate this to employees. They should also inform staff that input content may be processed by subcontractors.
Start by thinking back — have you ever pasted a password or API key into Notion AI?
この記事は AI Friends からのクロスポストです。