CVE-2026-65975
MediumCVSS 6.5Exploitation Probability (EPSS)
Low risk11th percentile - higher than 11% of all known CVEs
Summary
In Pydantic AI before versions 1.107.1 and 2.5.0, UI adapters (AG-UI and Vercel AI) incorrectly strip unresolved tool calls from message history. A client can cause a registered server tool to run with client-supplied arguments, bypassing the model and its guardrails.
Risk Assessment
The risk is remote execution of server tools with arbitrary arguments, potentially leading to unauthorized actions, especially in applications relying on pre-model hooks for access control.
Recommendation
Update Pydantic AI to version 1.107.1 or 2.5.0 depending on your branch.
Other vulnerabilities in Pydantic AI
See all- CVE-2026-54249Medium
In Pydantic AI before versions 1.106.0 and 2.0.0b6, the UI adapter (e.g., Vercel AI) does not validate UploadedFile references. A client can reference arbitrary files in the model provider or cloud storage account using server-side identity, potentially leading to unauthorized data access.
- CVE-2026-46678Medium
In Pydantic AI before version 1.99.0, when an application enables force_download='allow-local' for a URL, the cloud metadata blocklist can be bypassed by encoding the metadata IP in IPv6 form (e.g., IPv4-mapped IPv6), potentially exposing cloud IAM short-term credentials.
- CVE-2026-48782Medium
In versions 1.56.0 through 1.101.0 and 2.0.0b1 and 2.0.0b2 of the Pydantic AI framework, the cloud metadata blocklist could be bypassed by encoding the metadata IP in an IPv6 transition form, leading to the exposure of short-term IAM credentials in the cloud. The previous fix did not decode all IPv6 transition forms, resulting in a security vulnerability.
- CVE-2026-25580High
A Server-Side Request Forgery (SSRF) vulnerability exists in Pydantic AI's URL download functionality from version 0.0.26 to before 1.56.0. Attackers can include malicious URLs in message history, causing the server to make HTTP requests to internal network resources.
- CVE-2026-25640High
In Pydantic AI versions 1.34.0 through 1.50.0, a path traversal vulnerability in the web UI allows an attacker to craft a malicious URL that fetches and executes arbitrary JavaScript from an external source on the same CDN instead of the legitimate chat UI package. This affects applications using Agent.to_web or clai web to serve a chat interface.
Original NVD description (English source)
Pydantic AI is a Python agent framework for building applications and workflows with Generative AI. In versions 1.88.0 up to but not including 1.107.1 and 2.0.0b1 up to but not including 2.5.0, the UI adapters (AG-UI via Agent.to_ag_ui()/AGUIAdapter, and Vercel AI via VercelAIAdapter) use sanitize_messages to strip unresolved ("dangling") client-submitted tool calls from untrusted message history before it reaches the agent, a defense-in-depth default that prevents the agent from executing tool calls the model never emitted. However, the strip anchored to a message index computed before sanitization ran, so when a trailing client message sanitized to empty and was dropped (for example a client system message under the default manage_system_prompt='server'), a preceding assistant response carrying an unresolved tool call became the new tail and was dispatched without inspection. As a result, a remote client could cause a registered, non-approval server tool to run with client-supplied arguments rather than arguments the model produced. The impact is bounded by what the affected tools do and is most significant for applications that gate tool execution in a model-request hook (before_model_request / after_model_request), since a forged call skips the model turn and bypasses that guardrail; approval-gated tools (requires_approval=True) are not auto-executed by this path. This issue has been fixed in versions 1.107.1 and 2.5.0.