Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentAI infrastructure

LightLLM: four open pickle services let anyone on the network run code

CriticalModelTC

Mingkai Yu and Jiajia Liu, with Jiapeng Li on three of the reports, found four services of the same kind in LightLLM, an LLM inference and serving framework: RPyC services that accept connections with no authentication and with pickle deserialization enabled. VulnCheck published them as CVE-2026-96560 (the KV-transfer control channel started with --pd_trans_mode nccl, September 23), CVE-2026-103040 (the router profiler started with --enable_profiling, September 29), CVE-2026-103041 (the embed cache of multimodal deployments, September 29) and CVE-2026-103395 (visual-only deployments, September 30), each rated 9.3 under CVSS 4.0 for LightLLM 1.2.0 and earlier. Anyone who can reach one of these services can send a crafted object and run code as the LightLLM service account. When we checked on October 4, 1.2.0 was still the latest release and the GitHub issue for the visual-only flaw was open with no linked fix.

Replay the attack step by step in the Learning Center

How it happened

  1. LightLLM's components talk to each other over RPyC; four of these services accept connections with no authentication and allow pickle: the KV-transfer control channel (--pd_trans_mode nccl), the router profiler (--enable_profiling), the embed cache of multimodal deployments and the visual-only server.
  2. The embed cache and the visual-only server listen on all interfaces.
  3. A pickled object with a crafted __reduce__ method runs the attacker's code when the service unpacks it, for example through the visual server's remote_infer_images method.

Impact

Full control of the inference node and of everything its service account can reach, including models and GPU hosts.

The fix

No patched release yet: keep every LightLLM RPyC port off networks others can reach (bind it to localhost or firewall it), leave --enable_profiling off where it is not needed, and watch the LightLLM repository for a fix.

Lessons

  • Internal services between AI components are still network services; give them authentication or keep them local.
  • Pickle over the network is code execution by design.

Sources