LiteLLM, a proxy server for calling LLM APIs, prior to version 1.83.0, when configured with JWT authentication enabled, uses the first 20 characters of the JWT token as a cache key for OIDC userinfo. Because JWT headers produced by the same signing algorithm have identical first 20 characters, an unauthenticated attacker can craft a token that matches the cache key of a legitimate user. This allows the attacker to bypass authentication and assume the identity and permissions of that user. The vulnerability is classified as CWE-287 (Improper Authentication) and has a CVSS 4.0 score of 9.4 (critical). The issue is resolved in LiteLLM version 1.83.0.

An unauthenticated attacker can impersonate legitimate users by exploiting the cache key collision in JWT tokens, gaining unauthorized access to the victim's identity and permissions within affected LiteLLM deployments. This can lead to privilege escalation and unauthorized actions within the system. The impact is critical due to the ease of exploitation and the high level of access gained.

Upgrade affected LiteLLM deployments to version 1.83.0 or later, where this vulnerability is fixed. Since the issue only affects instances with JWT/OIDC authentication enabled (which is not enabled by default), disabling JWT authentication can also mitigate the risk. Patch status is not explicitly stated in the vendor advisory, but the fix is included in version 1.83.0. No cloud service remediation applies as this is not a cloud service.