LangSmith SDK Output Redaction Bypass Vulnerability in JavaScript and Python SDKs

A vulnerability exists in the LangSmith Client SDKs for JavaScript and Python, prior to the patched versions 0.5.19 and 0.7.31 respectively. The issue arises because the SDKs' output redaction controls do not apply to streaming token events. When a large language model (LLM) run generates streaming output, each token chunk is recorded as a new_token event with the raw token value, bypassing the redaction pipeline. This oversight can lead to sensitive LLM output being inadvertently stored in LangSmith, as the streamed content leaks through run events.

Impact

Exploitation of this vulnerability causes sensitive streamed content from LLM runs to be leaked via run events, undermining the output redaction controls that users may rely on to protect such information.

Remediation

Users can upgrade to version 0.5.19 of the JavaScript SDK or version 0.7.31 of the Python SDK to address this vulnerability.