Cyber & AI intelligence
Wasteland.
Briefs indexed2938
Issues30
Published Mondays07:30 CT
⚡ Active KEV CVE-2026-103040 2026-09-29

LightLLM Router Profiler Exposes Unauthenticated RPyC Service to Remote Code Execution (CVE-2026-103040)

"CVE-2026-103040 is a critical (CVSS 9.8) unauthenticated remote code execution flaw in LightLLM through version 1.2.0. It is caused by pickle deserialization in the router profiler's RPyC service."

CVE-2026-103040 is a critical (CVSS 9.8) unauthenticated remote code execution flaw in LightLLM through version 1.2.0. It is caused by pickle deserialization in the router profiler's RPyC service.

What Is It

LightLLM is ModelTC's LLM inference framework. It has a remote code execution vulnerability in its router profiler service. The vulnerable service only runs when LightLLM is started with the --enable_profiling flag. In that mode, the profiler exposes an RPyC server with no authentication and with pickle deserialization enabled. An attacker can run arbitrary code by sending crafted serialized objects to the profiler command queue.

The weakness is classified as CWE-502 (Deserialization of Untrusted Data). VulnCheck disclosed the issue and it was published to NVD on 2026-09-29. NVD currently lists its status as "Deferred."

Why It Matters

VulnCheck scores this 9.8 CRITICAL under CVSS 3.1 (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H) and 9.3 CRITICAL under CVSS 4.0. That means:

When the profiler is on, an attacker who can reach the RPyC service may be able to run code on the host serving the model.

Exploitation status: There is no CISA KEV entry for this CVE. KEV does not confirm active exploitation, and the CVSS 4.0 exploit maturity field is "Not Defined."

What's Vulnerable

The vulnerable code is in lightllm/server/router/profiler_service.py in the v1.2.0 tag (lines 32–34 and 46–50).

Patch Status

The source data does not name a fixed version. The NVD record lists every version through 1.2.0 as affected and gives no patched release. Neither CISA KEV nor NVD specifies a required action.

Operators should check the upstream GitHub issue (#1597) and the VulnCheck advisory for remediation updates. They should also find any LightLLM instances running with --enable_profiling, since the vulnerable service is only exposed in that configuration.

Sources