CVE-2026-133505: Tool result data shaped like a file is rebuilt as a file download or provider file reference when message history is reloaded
When message history is serialized and loaded again, a tool result that contains a JSON object in the serialized shape of one of Pydantic AI's file types is rebuilt as that file. A third party who controls data a tool returns, such as a record that an MCP server or a function tool returns from an external system, can use this to make the application download a URL on its local network, or make the model provider read an s3:// object or uploaded file with the application's credentials. Applications are affected when tool results can contain third-party data, message history containing those results is serialized and loaded again (stored history, history a UI adapter holds on the server, or Temporal, DBOS or Prefect durable execution), and the model is one Pydantic AI downloads files for, one that honors forcedownload, or one whose provider resolves non-HTTP file URLs or uploaded files. History that a client submits through the Vercel AI or AG-UI adapter in the default mode is not affected. The issue is fixed in pydantic-ai and pydantic-ai-slim 1.107.10 and 2.66.0.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
pydantic-aito a version that resolves this vulnerability.Fixed in 1.107.10 - Upgrade
Upgrade
pydantic-aito a version that resolves this vulnerability.Fixed in 2.66.0 - Upgrade
Upgrade
pydantic-ai-slimto a version that resolves this vulnerability.Fixed in 1.107.10 - Upgrade
Upgrade
pydantic-ai-slimto a version that resolves this vulnerability.Fixed in 2.66.0
Event History
Frequently Asked Questions
Which deployments are exposed to this issue?
An application is affected only when third-party data can appear in tool results, the resulting message history is serialized and later reloaded, and the selected model/provider handles the reconstructed file in a relevant way. Relevant handling includes Pydantic AI downloading files for the model, honoring force_download, or a provider resolving non-HTTP file URLs or uploaded files.
What must an attacker control to exploit it?
The attacker needs control over data returned by a tool, such as a record returned by an MCP server or a function tool backed by an external system. They must cause that data to contain a JSON object matching the serialized shape of a Pydantic AI file type, which is then present in history that is later reloaded.
What impact can successful exploitation have?
The crafted result can cause the application to download a URL reachable on its local network. It can also cause the model provider to read an s3:// object or an uploaded file using the application's credentials.
Are client-submitted histories through the Vercel AI or AG-UI adapter affected by default?
No. History submitted by a client through the Vercel AI or AG-UI adapter in its default mode is not affected.
Which versions contain the fix?
The issue is fixed in pydantic-ai and pydantic-ai-slim versions 1.107.10 and 2.66.0.