CVE-2026-63764 is a server-side request forgery (SSRF) vulnerability in InternLM's lmdeploy product affecting versions through 0.14.0. The flaw exists in the _load_http_url function within connection.py, where the private-IP guard performs validation only on the initial URL but fails to re-validate hosts after HTTP redirects. An attacker can exploit this by submitting a crafted image_url to the chat completions endpoint that points to an attacker-controlled server which issues an HTTP redirect to a private IP or cloud metadata endpoint. The server follows the redirect, potentially exposing internal service content through the model pipeline. The vulnerability is unauthenticated and has a CVSS 4.0 score of 7.7 (high severity). No official patch or remediation guidance is currently provided, but a fix was committed in a specific code commit (03c3130).

An unauthenticated attacker can exploit this SSRF vulnerability to cause the server to make HTTP requests to internal or cloud metadata endpoints by leveraging HTTP redirects. This can lead to exposure of sensitive internal service content through the model pipeline, potentially compromising internal network information or cloud instance metadata. The vulnerability has a high severity rating (CVSS 7.7) due to its ability to bypass private IP validation via redirects.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. The vulnerability is fixed in commit 03c3130, so applying the fix from that commit or upgrading to a version including this fix is recommended once available. Until then, restrict or monitor outbound HTTP requests from the affected service to prevent SSRF exploitation. Avoid using untrusted URLs in the image_url parameter of the chat completions endpoint.