CVE-2026-54249: Pydantic AI: Unvalidated UploadedFile references in UI adapters allow server-side file access using the application's credentials

UI adapters rebuild file parts from message history that the browser sends. File URL parts are checked against a scheme allowlist. UploadedFile parts are not. They point at a file by provider file ID or by a storage URI like s3://... or gs://....

The model provider resolves such a reference with the server-side identity: an IAM role, a service account or the provider API key. The client's own rights do not matter. A client that puts a chosen UploadedFile into its history makes the application's credentials fetch that object. The file content reaches the model, and so the client, through the model's answer. The objects can belong to the application's own account or to other tenants.

The attacker must know a valid identifier. The advisory says these are not always unguessable, depending on how the application names objects. That is why the CVSS vector has high complexity.

Only applications that pass untrusted, client-submitted message history to an agent through a UI adapter are exposed. An agent that builds its own messages server-side is not.

PyPI packages pydantic-ai and pydantic-ai-slim, versions >= 1.65.0, < 1.106.0 and >= 2.0.0b1, < 2.0.0b6. You are exposed if you use a UI adapter (the advisory names the Vercel AI adapter) and accept message history from clients. Check the versions:

pip show pydantic-ai pydantic-ai-slim

Upgrade to 1.106.0 on the 1.x line, or 2.0.0b6 on the 2.x beta line:

pip install --upgrade "pydantic-ai-slim>=1.106.0"

Both releases validate UploadedFile references on client-submitted messages the same way file URLs are validated.

If you cannot upgrade, the advisory gives two workarounds: do not pass untrusted client-submitted message history to the agent, or strip UploadedFile parts from incoming messages before the agent runs.

Review the message payloads your UI endpoint received for UploadedFile parts, particularly ones with s3:// or gs:// URIs or file IDs that your application did not issue. Compare them with the files you actually created. On the cloud side, check object-access logs (for example S3 or GCS audit logs) for reads by the application's role on objects that do not belong to the requesting user or tenant. Check the model provider's file-access logs for the same.

Summary

A client that can submit message history to a Pydantic AI UI adapter can reference arbitrary files in the application's model-provider or cloud-storage account. The server forwards the reference to the model provider, which fetches it using the server's own credentials, allowing the client to read files it should not have access to.

Details

UI adapters reconstruct file parts from client-submitted message history and forward them to the model provider. File URL parts are validated against a scheme allowlist before being forwarded, but UploadedFile references — which point to a file by provider file ID or cloud-storage URI (e.g. s3://…, gs://…) — were forwarded without validation.

Because the provider resolves an UploadedFile using the server-side identity (IAM role, service account, or provider API key) rather than the client's, a client that crafts message history containing an attacker-chosen UploadedFile can cause the server to read objects belonging to its own account or to other tenants, given a referenceable identifier.

Impact

Applications that pass untrusted client-submitted message history to an agent through a UI adapter (such as the Vercel AI adapter). Exploitation requires the attacker to reference a valid file identifier; depending on how the application names objects, such identifiers are not always unguessable.

Patches

Upgrade to 1.106.0 (1.x) or 2.0.0b6 (the 2.x beta line), which validate UploadedFile references on client-submitted messages the same way file URLs are validated.

Workarounds

If users cannot upgrade, do not pass untrusted client-submitted message history to the agent, or strip UploadedFile parts from incoming messages before running the agent.