Agent incident database
Notion AI agent leaked client data via its web search tool
19 Sept 2025Notion · Notion 3.0 AI agentsdata exfiltrationdisclosed vulnerability
CodeIntegrity showed that a PDF with hidden text could make Notion's new AI agent collect confidential client data and send it to an outside server. The data left through the agent's built-in web search tool.
How it unfolded
Each step on its own looked like ordinary agent work. The harm came from the sequence.
- 1
Read a PDF from an outside source containing hidden instructions
- 2
Model followed the hidden instructions
- 3
Read confidential client names, companies and revenue figures from the workspace
- 4
Called the web search tool with a URL carrying that data to the attacker's domain
harm
Which policy would have stopped this?
In our reading of the reports, these policies from the open library stop the harmful step before it runs.
- Block external send after sensitive data read: The web search call carried client data the agent had read earlier in the task, so it is blocked before it runs.
- Restrict outbound to approved domains: The search went to an address outside the approved domains.
Open Block external send after sensitive data read in the playground and build the path yourself.
This is our reading of public reporting. We haven't run it against the original systems.
What happened
Shortly after Notion launched AI agents in Notion 3.0, Abi Raghuram of CodeIntegrity published a demonstration using Claude Sonnet 4. A PDF styled as a customer feedback report contained white-on-white instructions disguised as an internal work task. When a user asked the agent to summarise the report, it read the workspace's client list, combined client names, companies and ARR figures into one string, and appended it to a URL on an attacker-controlled domain. It then called Notion's functions.search web tool with that URL as the query, which delivered the data to the attacker's server logs.
Outcome
Notion said it upgraded its prompt-injection detection, including for file attachments, added a user approval step for suspicious or model-generated links, and gave administrators controls to restrict or disable agents' web access.
Sources
- Notion AI Security: Web Search Exfiltration Risk, CodeIntegrity, 2025-09-19
- The Hidden Risk in Notion 3.0 AI Agents: Web Search Tool Abuse for Data Exfiltration, Simon Willison's Weblog, 2025-09-19
- Notion AI agents get security update after potential data leak, The Decoder, 2025-09-21
A summary of public reporting. The step-by-step timeline is our reading of those sources; where they differ, we say so above.
Your agents
Could this happen to yours?
For developers
Build a path yourself
Recreate a sequence in the playground and see how policies treat it.