LightLLM: Denial of service via unsafe deserialization in /generate endpoint

Summary

LightLLM through version 1.2.0 contains a denial-of-service vulnerability caused by unsafe deserialization (CWE-502). An unauthenticated attacker can send a single request to /generate with a crafted internal pd_kv_trans_params field, which crashes the inference worker thread. The process and the /health endpoint keep running, but all later requests hang until the service is restarted.

Affected

• LightLLM 0 to 1.2.0 (all versions up to and including 1.2.0)

Fixed in

Unknown. No fixed version has been published as of the sources reviewed.

Workaround

• Restrict access to /generate to trusted clients, for example behind an authenticating reverse proxy or within an internal network.
• Check whether the pd_kv_trans_params field can be sent from untrusted sources and, if possible, filter it at the proxy after verifying that PD mode still works.
• Monitor inference availability with a test generation request, not only /health.

Exploitation

• Exploitation status: no known exploitation in the sources reviewed; not listed as known exploited.
• CISA SSVC: exploitation status poc, automatable yes, technical impact partial.
• VulDB states that no exploit is available, but the GitHub issue contains a detailed root-cause description and reproduction steps.

Our recommendation

• Identify all LightLLM instances and their versions and assume that versions up to 1.2.0 are affected.
• Check whether /generate, PD mode, or other management endpoints are reachable from outside the trusted network and restrict access now.
• Monitor the GitHub issue and NVD for a fixed release and update as soon as one is available.
• Add a monitoring check that sends a test generation request, so that a hung inference loop is detected.

Public exploits iAttack code is publicly available (PoC on GitHub, Metasploit or Exploit-DB).

PoC-in-GitHub, Metasploit and Exploit-DB list nothing for this CVE.

Research sources

• LightLLM issue #1629 (reporter's root-cause analysis and PoC) github.com
• VulnCheck advisory: LightLLM through 1.2.0 unauthenticated DoS www.vulncheck.com
• LightLLM v1.2.0 api_lightllm.py github.com
• LightLLM v1.2.0 sampling_params.py github.com
• LightLLM v1.2.0 infer_batch.py github.com
• VulDB entry CVE-2026-111160 vuldb.com
• HOL CVE record CVE-2026-111160 hol.org
• OpenCVE CVE-2026-111160 app.opencve.io

Researched and written up in English and German by an AI (Claude by Anthropic) based on NVD, CISA, EPSS and the sources listed. The vendor information is authoritative. Researched on 1 Nov 2026, 02:26.