# CVE-2026-103042: LightLLM memory exhaustion in NCCL control channel

Source: https://junglewise.ai/threats/cve-2026-103042-lightllm-through-1-2-0-contains-a-memory-exhaustion  
Publisher: Junglewise Threat Intelligence, https://junglewise.ai/threats  
Published: 2026-09-29. Updated: 2026-09-29  
License: CC BY 4.0 with a link back

CVE-2026-103042 · Severity: high · CVSS 7.5 · Published 2026-09-29

LightLLM is a framework for running large language models on distributed GPU clusters. When configured to use NCCL for cross-node communication (--pd_trans_mode nccl), it exposes an unauthenticated control service that allows any network attacker to exhaust worker memory by storing unbounded data, crashing the entire inference node and interrupting service.

Technologies: [ModelTC LightLLM](https://junglewise.ai/threats/technologies/lightllm). Vendors: [ModelTC](https://junglewise.ai/threats/vendors/modeltc).

## Quick answers

**Is CVE-2026-103042 exploited in the wild?** No. Junglewise Threat Intelligence has no report of CVE-2026-103042 being exploited in the wild.

**How severe is CVE-2026-103042?** High severity, with a CVSS score of 7.5, published 29 September 2026.

**Which products and versions are affected?** ModelTC LightLLM through 1.2.0.

**Is there a fix for CVE-2026-103042?** No fixed version is recorded yet. Check the vendor's advisory in the references below. If you run it, Junglewise drafts a mitigation that fits your setup and applies it once you approve, with senior security engineers behind it.

**How is CVE-2026-103042 attacked?** It is attacked over the network, in a software library.

**Am I affected by CVE-2026-103042?** Check whether you run ModelTC LightLLM at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

## Technical details

The NCCL control channel in LightLLM's KV-transfer workers runs an unauthenticated RPyC service on ports 30000-40000 with the exposed_set_value method. This method stores caller-supplied key-value pairs in an unbounded Python dictionary with no size limits, entry limits, or expiration, allowing attackers with network access to exhaust memory. When the worker process is OOM-killed, a watchdog escalates the failure by forcibly terminating the entire router process tree with SIGKILL.

## Affected products

- ModelTC LightLLM through 1.2.0

## Timeline

- 2026-09-29: disclosed

## References

- [https://github.com/ModelTC/LightLLM](https://github.com/ModelTC/LightLLM)
- [https://github.com/ModelTC/LightLLM/issues/1595](https://github.com/ModelTC/LightLLM/issues/1595)

## Related threats

- [LightLLM denial of service via pd_kv_trans_params unpickling](https://junglewise.ai/threats/cve-2026-111160-lightllm-through-1-2-0-contains-a-denial-of-service)
- [LightLLM visual_only unauthenticated RCE via RPyC pickle deserialization](https://junglewise.ai/threats/cve-2026-103395-lightllm-through-1-2-0-visual-only-deployments-expose-an)
- [LightLLM authentication bypass in reinforcement learning control API](https://junglewise.ai/threats/cve-2026-103270-lightllm-through-1-2-0-mounts-reinforcement-learning-control)
- [LightLLM server-side request forgery in multimodal endpoints](https://junglewise.ai/threats/cve-2026-103243-lightllm-through-1-2-0-fails-to-validate-image-url-and-audio-url)
- [LightLLM multimodal pickle deserialization in RPyC cache service](https://junglewise.ai/threats/cve-2026-103041-lightllm-through-1-2-0-multimodal-deployments-expose-an)