ModelTC LightLLM: Missing authentication on RL control routes

Summary

LightLLM through 1.2.0 mounts reinforcement learning control routes on the public HTTP API without any authentication check. On deployments started with --enable_rl, an unauthenticated attacker with network access to the inference port can pause generation, abort in-flight requests, flush the prompt cache and wedge inference workers. The impact is limited to availability; confidentiality and integrity are not affected according to the CVSS v4.0 vector.

Affected

• LightLLM 0 up to and including 1.2.0
• Primarily deployments started with --enable_rl
• The reporter states the routes are also mounted without --enable_rl, where they return errors instead of 404 (not yet independently confirmed)

Fixed in

Unknown. A pull request (#1662) proposing authentication for the RL routes exists, but its merge and release status could not be confirmed.

Workaround

• Unknown. No confirmed patch or vendor-approved workaround.
• Keep the inference port off the internet and allow only known internal hosts.
• Place an authenticating reverse proxy in front of the service and block the RL endpoints (/pause_generation, /abort_request, /flush_cache, /init_weights_update_group and others).
• Test any workaround in a non-production environment first.

Exploitation

• CISA SSVC: exploitation 'poc', automatable 'yes'.
• Not listed in CISA KEV.
• Public PoC availability is disputed between sources and could not be confirmed.
• Treat exploitation as plausible and unconfirmed.

Our recommendation

• Find out whether LightLLM is in use, including in AI or ML pilot projects.
• Check whether its ports are reachable from outside and restrict access.
• Monitor logs for calls to the RL endpoints named in the CVE.
• Track PR #1662 and LightLLM releases for a confirmed fixed version, then update.
• Also review CVE-2026-103040, a reported remote code execution issue in LightLLM up to 1.2.0 (router profiler, when profiling is enabled).

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

• CVE-2026-103270 Issue #1609 (ModelTC/LightLLM) github.com
• VulnCheck advisory: LightLLM missing authentication on RL control routes www.vulncheck.com
• LightLLM api_http.py at v1.2.0 github.com
• LightLLM api_http_rl.py at v1.2.0 github.com
• LightLLM GitHub repository github.com
• CVE-2026-103270 on OpenCVE app.opencve.io
• OSV entry CVE-2026-103270 osv.dev
• Pull request #1662 (ModelTC/LightLLM) github.com

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, 06:37.