Loading vulnerability…
CVE-2026-33626 is a server-side request forgery vulnerability in LMDeploy versions before 0.12.3. The vision-language image-loading functionality dereferences attacker-supplied image URLs without validating resolved hosts or blocking loopback, private, and link-local address ranges. An attacker can cause a vulnerable LMDeploy inference server to issue HTTP requests to arbitrary external or internal destinations, including cloud metadata services.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
Patch, then assume compromise.
1 valid exploit after Mallory filtered fakes, detection scripts, and README-only repos.
This repository is a local Docker-based SSRF lab for CVE-2026-33626 affecting InternLM LMDeploy vision-language image loading. It is a real proof-of-concept repository, not just documentation. The core exploit capability is to supply an attacker-controlled URL to a FastAPI wrapper endpoint (/probe), which then calls lmdeploy.vl.load_image(url). In the vulnerable build (LMDeploy 0.12.0), this causes a server-side fetch of internal/private resources, demonstrating SSRF. In the patched build (LMDeploy 0.12.3), the same request is rejected with a security error. Repository structure: docker-compose.yml orchestrates three services: vuln, patched, and internal. The vuln/Dockerfile and patched/Dockerfile each build a minimal FastAPI/uvicorn wrapper around LMDeploy and expose endpoints /, /version, and /probe. The internal/server.py file implements a simple HTTP canary service on port 9000 that serves /private.png and records requests in memory, exposing /hits and /reset for verification. The poc/poc.py script is the main operator entry point; it resets the canary, probes both vulnerable and patched services, and compares whether the internal service was contacted. Exploit flow: poc/poc.py sends a GET request to /probe?url=<encoded target> on the vulnerable service. The vulnerable wrapper passes the supplied URL into lmdeploy.vl.load_image(), which fetches and decodes the remote image. Success is confirmed when the internal canary logs a hit to /private.png. The same flow against the patched service should fail before any internal request occurs. This is an operational SSRF PoC with a fixed, benign payload URL rather than a customizable post-exploitation payload. It demonstrates internal network reachability and server-side URL fetching, but does not include RCE, persistence, credential theft, or lateral movement logic.
Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.
Vendor-confirmed product mapping. Mallory continuously reconciles this list against your asset inventory.
56 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
An SSRF vulnerability in LMDeploy's load_image() function that can expose cloud metadata services and internal networks. The article contrasts it with the primary CVEs because its scope is unchanged and its integrity impact is none.
Server-side request forgery vulnerability in LMDeploy, mentioned as one of the additional cases in the honeypot methodology results.
A vulnerability in LMDeploy cited only as an example of rapid post-disclosure exploitation in another ecosystem.
An SSRF vulnerability in LMDeploy that was exploited extremely rapidly after disclosure, illustrating AI-compressed time-to-exploitation.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.