CVE-2026-33626: High-Severity SSRF Exploited in the Wild to Hijack AI Inference Engines

Sysdig Threat Research disclosed CVE-2026-33626, a high-severity SSRF in LMDeploy (<= v0.12.2) that allows attackers to use the image URL loader as a generic HTTP SSRF primitive; exploitation was observed within 12 hours and 31 minutes against honeypots, with attackers probing AWS IMDS for credentials, Redis, external egress, administrative APIs, and a distributed-serving kill-switch. Recommended mitigations include immediate patching to v0.12.3 or later, enforcing IMDSv2 with required httpTokens, strict VPC egress controls, and monitoring outbound/internal connections from inference processes.