CVE-2026-133505
Pydantic AI could rebuild untrusted tool data as privileged file references
pydantic · Pydantic AI / pydantic-ai-slim
First documented disclosure · Review checkpoint
Summary
Reloaded tool results could cause private-network downloads or provider-side reads using application credentials.
Severity and exploitation evidence
Medium · 5.9/10 · CVSS 3.1. Severity source.
Exploitation not established: Reviewed public sources describe the weakness, a correction or a controlled demonstration; malicious exploitation against users is not established.
Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.
How the weakness works
File-shaped JSON data becomes a typed file when serialized message history is reloaded.
Conditions required
Third-party content reaches tool results, history is serialized/reloaded, and the model or provider supports the relevant file handling.
Potential impact
Internal responses may reach the model provider; provider-readable objects can be fetched with application credentials. Attacker retrieval requires a subsequent disclosure path.
Affected versions and configurations
1.50.0–1.107.9 and 2.0.0b1–2.65.0 for both packages.
Fix and mitigation status
1.107.10 and 2.66.0.
Practical next steps
Upgrade; sanitize saved history. Finish or discard old durable/harness runs containing untrusted tool data, or sanitize before continuing them. Do not enable legacy file rebuilding for untrusted history.
Operating lesson
Persist a trusted type marker separately from data controlled by tool-result authors.
Evidence and limits
Default client-submitted Vercel AI and AG-UI history is sanitized and unaffected. allow-local still blocks cloud-metadata addresses. Old durable journals and harness snapshots retain earlier decoding behavior to preserve in-flight runs, so an upgrade alone does not sanitize those runs. Public advisory publication date is verified below. CNA publication is 2026-11-20; this is distinct from the earliest reviewed public source and correction dates. Pydantic assessments: CVSS 3.1 5.9.
Sources
- Pydantic CVE authority record
Primary source ·
- Pydantic AI original advisory
Primary source ·