CVE-2026-127222
7.5HighPublished 18 Nov 2026 · NVD status: Deferred · AI triage by gemini-3.1-flash-lite
High
0CVSS v3.110
- Attack vector
- Network
- Complexity
- Low
- Privileges
- None
- User action
- None
- Impact
- High impact on integrity
- Weakness
- CWE-862
What an attacker gets
Unauthenticated network attackers can load arbitrary LoRA adapters to manipulate model output. This can be chained with a deserialization bypass to achieve remote code execution as root.
Root cause
The /load_lora_adapter_from_tensors endpoint is missing the required admin authorization decorator.
Affected versions
SGLang through 0.5.23
CVSS v3.1 · CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:H/A:N
Fix & remediation
Upgrade to a version containing commit 7b5fb39. Until patched, block network access to the SGLang API endpoint /load_lora_adapter_from_tensors at the firewall or reverse proxy level.
$ pip install --upgrade sglangAI-suggested. Run it in staging first and check it against the vendor advisory. Why?
NVD description
SGLang through 0.5.23, fixed in commit 7b5fb39, contains a missing authorization vulnerability in the POST /load_lora_adapter_from_tensors HTTP endpoint, which lacks the @auth_level(AuthLevel.ADMIN_OPTIONAL) decorator applied to its sibling admin endpoints, allowing unauthenticated network-reachable attackers to invoke it without an API key. Attackers can load a LoRA adapter from attacker-supplied serialized tensors to manipulate the served model, and when combined with a separately documented deserialization bypass, achieve unauthenticated remote code execution as root.
References
- https://github.com/rahulreddykarne/CVE-2026-127222-sglang
- https://github.com/sgl-project/sglang/commit/7b5fb39bb952da5f6da3bb7fe73c671cf449b6fa
- https://github.com/sgl-project/sglang/pull/49896
- https://rahulkarne.com/#/cve/CVE-2026-127222
- https://www.vulncheck.com/advisories/sglang-missing-authorization-via-load-lora-adapter-from-tensors