LMDeploy Vulnerability Exposes LLM Servers to SSRF Attacks

The National Vulnerability Database has identified a critical Server-Side Request Forgery (SSRF) vulnerability in LMDeploy, a popular toolkit for large language model (LLM) deployment. Affecting versions prior to 0.12.3, the flaw resides within the vision-language module. Specifically, the load_image() function fails to validate arbitrary URLs, allowing unauthenticated attackers to fetch content from internal networks and cloud metadata services.

This SSRF vulnerability, rated HIGH with a CVSS score of 7.5, presents a significant risk. Attackers can exploit this weakness to pivot within an organization’s infrastructure, access sensitive cloud credentials, or exfiltrate internal data by tricking the LMDeploy server into making requests to attacker-controlled or sensitive internal endpoints.

Organizations utilizing LMDeploy must immediately upgrade to version 0.12.3 or later. Defenders should also consider implementing network segmentation and egress filtering to limit the potential blast radius of any successful SSRF exploitation, even on patched systems.