LangGraph load operations deserialize checkpoint data that is encoded with msgpack, automatically reconstructing Python objects during the process. This uncontrolled deserialization is a classic instance of unsafe object reconstruction, listed under CWE‑502. Because a crafted payload can influence the reconstruction logic, an attacker who can supply malicious data during checkpoint loading may be able to execute arbitrary code or otherwise compromise the running process.

The vulnerability exists in LangChain AI’s LangGraph package version 1.0.9 and earlier. The affected component is the SQLite‑based checkpoint loader used by LangGraph, which stores checkpoints in a local SQLite database. Any deployment that relies on these older LangGraph versions and allows the checkpoint database to be written to by untrusted parties is exposed.

The CVSS score of 6.8 indicates moderate severity, but the EPSS score of less than 1% suggests that actual exploitation is currently unlikely. The vulnerability is not listed in the CISA KEV catalog, so no public exploits have been reported yet. Nevertheless, the attack requires the ability to alter the checkpoint database—either through a database compromise or privileged write access. If an adversary succeeds in modifying the stored checkpoint, they can inject a payload that triggers unsafe deserialization, potentially leading to remote code execution.