CVE-2026-33626 is an SSRF vulnerability in LMDeploy, an LLM inference toolkit, that was exploited within 12 hours and 31 minutes of its GitHub advisory publication. Sysdig's Threat Research Team captured the attack on a honeypot: over an 8-minute session, the attacker used the vision-language image URL loader as a generic HTTP probe to query AWS IMDS for IAM credentials, scan loopback ports (Redis 6379, MySQL 3306, HTTP 8080), perform OOB DNS exfiltration via requestrepo.com, enumerate the OpenAPI schema, and probe the unauthenticated /distserve/p2p_drop_connect endpoint to disrupt inference routing. No public PoC existed; the advisory text alone was sufficient to construct the exploit. Defenders are advised to upgrade to LMDeploy v0.12.3, enforce IMDSv2, restrict VPC egress from inference nodes, and use runtime detection rules (e.g., Falco) to catch IMDS access from containers.

10m read timeFrom webflow.sysdig.com
Post cover image
Table of contents
Exploitation timelineThe LMDeploy vulnerabilityThe three phases of LMDeploy exploitationWhat this means for defendersIndicators of CompromiseRuntime detectionRecommendationsConclusion
2 Impressions