What Happened?
A high-severity security flaw (CVE-2026-33626, CVSS score: 7.5) has been found in LMDeploy, an open-source toolkit for compressing, deploying, and serving large language models. The vulnerability is a Server-Side Request Forgery (SSRF) flaw in the toolkit's vision-language module. The load_image() function in lmdeploy/vl/utils.py fetches arbitrary URLs without validating internal or private IP addresses, allowing attackers to access cloud metadata services, internal networks, and other sensitive resources. The flaw was discovered by Orca Security researcher Igor Stepansky and was actively exploited in the wild within 13 hours of its public disclosure.
Am I Affected?
All versions of LMDeploy up to and including 0.12.0 that use vision language support are affected.
What Should You Do?
If you are running an affected version of LMDeploy with vision language support, treat this as urgent. Successful exploitation can allow attackers to steal cloud credentials, access internal services not exposed to the internet, scan internal networks, and establish lateral movement within your infrastructure. Review the official security advisory published by the project maintainers on GitHub (GHSA-6w67-hwm5-92mq) and apply any available patches or mitigations immediately.