AAK-LMDEPLOY-VL-SSRF-001
LMDeploy VL image loader fetches user-controlled URLs without allow-list
| Field | Value |
|---|---|
| Severity | HIGH |
| Category | TRANSPORT_SECURITY |
| Shipped | v0.3.6 |
| Scanner | agent_audit_kit/scanners/ssrf_redirect.py (extension) |
| CWE | CWE-918 |
| OWASP MCP | none |
| OWASP Agentic | ASI02 (Tool Misuse and Exploitation) |
| CVE | CVE-2026-33626 (GHSA-only at v0.3.6 cut — NVD enrichment pending) |
What it catches
Vision-language pipelines that call lmdeploy.serve.vl_engine.*
preprocessing helpers (preprocess_image_url, load_image(url=...),
encode_image(url=...)) without an SSRF guard. Same shape as
AAK-LANGCHAIN-SSRF-REDIR-001
but tied to the VL image loader.
Detection
- File must import
lmdeploy. - File must call a
lmdeploy.serve.vl_engine.VLEngine/preprocess_image_url(...)/load_image(url=...)/encode_image(url=...)shape. - Suppress if the same file has a known SSRF guard
(
validate_safe_url,is_safe_url,ensure_safe_url,ssrf_guard,ALLOWED_HOSTS,TrustedHostMiddleware).
Remediation
Wrap the URL with an SSRF guard before passing it to the VL loader. Resolve the hostname once, validate the resolved IP against an allow-list, and pin the resolved IP for the actual request. See AAK-SSRF-TOCTOU-001 for the canonical pinning recipe.
Note on metadata
NVD enrichment was pending at v0.3.6 cut (2026-04-26). The rule cites the GHSA index entry as the primary advisory. CVSS / CWE will tighten once NVD lands.
Registry reference
This section is generated from the rule registry (agent_audit_kit/rules/builtin.py) when the documentation is built, so it follows the scanner. The rest of the page is hand-written; where the two disagree, the registry is what runs.
- Severity
- HIGH
- Category
- Transport Security (
transport-security) - Scanner
agent_audit_kit/scanners/ssrf_redirect.py- Auto-fixable
- No.
- SARIF name
LmdeployVlSsrf
What it detects
A vision-language pipeline calls lmdeploy.serve.vl_engine.* (or framework-equivalent) preprocessing helpers with a URL argument that is not validated against an allow-list. CVE-2026-33626 (GHSA-only at time of v0.3.6 cut — NVD enrichment pending) documents this exact shape: an attacker submits an image URL that points at a private endpoint, the loader fetches it server-side, and the response is processed by the VL pipeline. Same SSRF class as AAK-LANGCHAIN-SSRF-REDIR-001 but tied to the VL image loader.
How to fix
Wrap the URL with the same SSRF guard you use for any other fetch: validate the resolved IP against an allow-list, disable redirects, and pin the resolved IP for the actual request. Upgrade lmdeploy to >= 0.12.3 and pin it: versions below 0.12.3 are affected (GHSA-6w67-hwm5-92mq). Exploitation was observed within roughly twelve hours of the advisory, so treat an unpinned or older install as urgent rather than routine.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-LMDEPLOY-VL-SSRF-001
Limitations
Looks in a Python file that mentions lmdeploy for load_image(url, encode_image(url, preprocess_image_url(, VLEngine() or the lmdeploy.serve.vl_engine.VLEngine path, and skips the file if it contains an SSRF guard name, ALLOWED_HOSTS or TrustedHostMiddleware. The argument has to begin with url: load_image(image_url) is not matched. A Python file over 1 MB is not read.
References
- CVE: CVE-2026-33626
- OWASP Agentic Top 10: ASI02 (Tool Misuse and Exploitation)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-NETWORK-04
- Incidents and advisories: GHSA-LMDEPLOY-VL-2026-04-25
- CSA AI Controls Matrix (AICM): IVS-04, AIS-08