The langsmith-sdk JavaScript and Python SDKs before versions 0.5.19 and 0.7.31 do not apply output redaction controls to streaming token events generated during LLM runs. Specifically, each chunk of streaming output is recorded as a new_token event containing the raw token value, which bypasses the redaction pipeline that only processes inputs and outputs fields. Consequently, sensitive streamed content can be leaked via run events despite redaction settings. The issue is fixed in versions 0.5.19 (JS) and 0.7.31 (Python).

Sensitive information generated by language model streaming outputs may be exposed to unauthorized actors because the redaction controls fail to sanitize streaming token events. This can lead to leakage of confidential or sensitive data stored within the LangSmith platform through run event logs. The vulnerability does not affect integrity or availability but impacts confidentiality.

Upgrade to langsmith-sdk JavaScript version 0.5.19 or later and Python version 0.7.31 or later. These versions fix the issue by applying output redaction controls to streaming token events. No other mitigation is indicated or required.