Description

Security researchers from Miggo Security have identified a critical vulnerability in LangSmith, tracked as CVE-2026-25750, which could enable session token theft and full account takeover. LangSmith is widely used to debug and monitor large language model (LLM) applications and processes billions of events daily, making the vulnerability particularly serious for organizations that rely on AI observability platforms. The issue originated from an insecure API configuration feature in LangSmith Studio. The platform allowed developers to define a flexible baseUrl parameter so frontend applications could send API requests to different backend services. However, earlier versions did not validate the destination domain. As a result, attackers could craft a malicious URL or webpage that redirected a victim’s browser to send API requests — including active session credentials — to an attacker-controlled server. If an authenticated user visited such a malicious page or a compromised website running hostile JavaScript, their browser could silently send the session token to the attacker. Because the attack leverages the victim’s active session, it does not require the user to enter credentials. Attackers who intercept the token have roughly five minutes before it expires, giving them enough time to hijack the account. Once inside, they could access AI trace histories, raw debugging data, proprietary system prompts, and potentially sensitive information such as internal code or customer data. LangChain addressed the vulnerability by enforcing a strict allowed-origins policy that requires domains to be pre-approved before they can be used as API base URLs. According to the official advisory published on January 7, 2026, there is no evidence of active exploitation. The issue was already fixed for LangSmith Cloud users in December 2025, while self-hosted deployments must upgrade to version 0.12.71 or later to remain protected.