The flaw resides in the add_text function of the Arena Side‑by‑Side View handler in FastChat. An attacker can manipulate input to cause the function to alter the normal control flow, producing an unintended sequence of operations. The vulnerability is identified as CWE‑670, indicating an incorrect or missing return value that can change the logic of the application. Because the defect propagates through remote API calls, an attacker can trigger the miscontrol flow from outside the host without local privileges. The remote nature and public availability of exploit code elevate the risk for any installation running a version of FastChat before 0.2.37.

FastChat, a conversational‑AI framework maintained by lm‑sys, is affected in all releases up to and including 0.2.36. Versions beyond 0.2.36 are presumed to incorporate the fix in commit 34eca62, but availability of a comprehensive patch that addresses all missed files has yet to be confirmed.

The CVSS score of 6.9 places the bug in the moderately severe range, yet its exploitation can have serious consequences because a misdirected control flow can lead to unauthorized operations within the application. With no EPSS score available, the exact likelihood of exploitation in the wild cannot be quantified, but the public release of exploit code means attackers have ready tools to target vulnerable deployments. Furthermore, the fault remains unpatched in three previously overlooked files, so a complete fix has not been rolled out. The combination of remote attack vector, public exploit, and incomplete patching makes this vulnerability a significant risk for entities running older FastChat instances.