CVE-2026-48775: LangChain LangGraph unsafe deserialization in JsonPlusSerializer

Executive brief

LangGraph is a library used to build stateful multi-agent applications with LLMs. A vulnerability in how it handles saved application states (checkpoints) could allow an attacker with write access to the database or storage layer to execute malicious code. This could lead to a full takeover of the application worker process and exposure of sensitive runtime secrets.

Technical details

The vulnerability exists in the `JsonPlusSerializer` component of LangGraph, which is used by default in persistent checkpointers like `PostgresSaver` and `SqliteSaver`. The serializer's deserialization path (CWE-502) could be manipulated to reconstruct arbitrary Python objects (CWE-913) from JSON payloads. An attacker who gains write access to the backing checkpoint store can inject malicious payloads that trigger arbitrary code execution when the application loads or resumes a checkpoint. The fix in version 4.1.1 restricts object reconstruction to a narrow set of safe default constructors.

Affected products

• langchain-ai langgraph-checkpoint < 4.1.1

Timeline

• 2026-05-22: disclosed
• 2026-06-16: advisory: NVD publication date
• 2026-06-25: patched: GitHub Advisory reviewed and updated

References

• https://api.github.com/users/pucagit
• https://github.com/pucagit
• https://api.github.com/users/pucagit/gists%7B/gist_id%7D
• https://api.github.com/users/pucagit/repos
• https://avatars.githubusercontent.com/u/93117777?v=4
• https://api.github.com/users/pucagit/events%7B/privacy%7D