Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessReach exposed NCCL control channel
-
DeliveryCall unauthenticated exposed_set_value RPC
-
ExploitStore unbounded key-value pairs
-
ExecutionExhaust KV-transfer worker memory
-
ImpactCrash worker and fail node
Vulnerability Assessment
| Weakness |
CWE
OWASP Top 10:2025
A06:2025 - Insecure Design
ASVS 5.0
V2 - Validation and Business Logic Resource allocation limits
MITRE ATT&CK
1 technique across 1 tactic: Impact The technique list with ATT&CK links and the kill-chain view is part of the Basic plan and up. |
| Exploitation Conditions | Requires the LightLLM server (through v1.2.0) to be started with the non-default |
| Impact | This is a genuine but bounded availability threat. The CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N/VA:H) describes a network-reachable, low-complexity, unauthenticated denial-of-service with high availability impact and zero confidentiality or integrity impact - consistent with the CWE-770 (unbounded resource allocation) memory-exhaustion described. No EPSS score, no CISA KEV listing, and no public exploit were provided, so the correct read is 'no public exploit identified at time of analysis.' The 8.7 score is driven almost entirely by the availability metric under the newer 4.0 scoring, which weights an unauthenticated remote DoS heavily; this is a real, if single-dimension, risk rather than an inflated critical. The exploitation is strictly conditional: it only exists when the server is launched with the non-default |
| Exploit ScenarioDerived | An attacker with network reach to a LightLLM node running in NCCL prefill-decode transfer mode repeatedly invokes the unauthenticated |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Sep 29, 2026
Memory exhaustion in LightLLM through 1.2.0 lets unauthenticated remote attackers crash KV-transfer worker processes when the server is started with the non-default --pd_trans_mode nccl flag. Under that mode the NCCL-based prefill-decode control channel exposes a set_value method that stores caller-supplied key-value pairs with no size or count limit, so repeated calls consume worker memory until the process dies and the node fails. Impact is availability-only (CVSS 4.0 base 8.7, VA:H; independently assessed as CVSS 3.1 AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H) with no data disclosure or modification; exploitation requires the opt-in NCCL transport mode plus network reachability to the control channel, and no public exploit has been identified at time of analysis.
RemediationAI
No vendor-released patch version is identified in the available intelligence at time of analysis; the supplied references point to the upstream source files at v1.2.0 and issue https://github.com/ModelTC/LightLLM/issues/1595 rather than to a patched release, so upgrade status should be confirmed against https://github.com/ModelTC/LightLLM and the VulnCheck advisory https://www.vulncheck.com/advisories/lightllm-through-1.2.0-unauthenticated-memory-exhaustion-via-nccl-control-channel-set-value before relying on a fixed build. The most effective immediate mitigation is to stop launching the server with --pd_trans_mode nccl and use the default (non-NCCL) PD transport, which removes the vulnerable code path entirely at the cost of the NCCL-specific KV-transfer performance profile. If NCCL transport is required, restrict network reachability of the NCCL control channel to the known prefill and decode peer nodes only using host firewalling or a network policy on the allocated port range, accepting that any misconfigured peer range will break disaggregated serving; place the channel behind a broker or sidecar that terminates the connection and enforces request size and rate limits, which adds a hop and modest latency but caps the number of stored entries per client. Run KV-transfer workers under a cgroup memory limit with supervised automatic restart and node-level health checks so an in-progress exhaustion degrades to a fast worker restart instead of a full node failure, understanding that this contains rather than prevents the crash, and monitor worker RSS growth and control-channel call rates for the sustained set_value pattern that characterizes exploitation.
Affected VersionsENISA
Affected Packages
Source: OSV (Open Source Vulnerabilities)
Affects your stack
Affects 1 package in euvd
Checking this CVE against your own stacks and servers is part of the Basic plan and up.
Technical ContextAI · show details
LightLLM is a high-throughput LLM inference and serving framework whose prefill-decode (PD) disaggregation mode splits prompt processing and token generation across separate worker pools, exchanging KV-cache blocks over a transport backend. CVE-2026-103042 concerns the NCCL transport variant (lightllm/server/router/model_infer/mode_backend/pd/nccl_kv_transporter.py, lines 416-420 and 474-485, with supporting logic in base_kv_move_manager.py lines 121-122 at tag v1.2.0), where the control channel exposes a set_value RPC that writes arbitrary key-value entries into a shared dictionary without enforcing any bound on key size, value size, or entry count. The root-cause class is CWE-770, Allocation of Resources Without Limits or Throttling: because the control channel is reachable over the network with no authentication (CVSS 4.0 vector AV:N/AC:L/AT:N/PR:N/UI:N, and PR:N/UI:N in the assessed v3.1 vector), an attacker can iterate the call and grow the worker's memory footprint until the process is OOM-killed, which in a disaggregated deployment also removes the node from the serving topology. The affected CPE cpe:2.3:a:modeltc:lightllm:*:*:*:*:*:*:*:* indicates the issue is tracked across all LightLLM releases up to and including 1.2.0 rather than a single build; there is no confidentiality or integrity component (VC:N/VI:N/SC:N/SI:N), so the impact is confined to worker availability and the cascading node failure described by the reporter (VulnCheck).
Know when Lightllm gets its next vulnerability
- Watch products and get an email when something new is published
- Check this CVE against your own stack and servers
- Exploit and PoC references, threat intelligence from 22 sources
No credit card · 7-day full trial
Severity by source
CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N
Unauthenticated network-reachable RPC with no user interaction and low complexity causes worker crash; availability-only impact, no confidentiality or integrity effect, scope unchanged.
Primary rating from Vendor (VulnCheck).
CVSS VectorVendor: VulnCheck
DescriptionCVE.org
LightLLM through 1.2.0 contains a memory exhaustion vulnerability in the NCCL control channel when started with --pd_trans_mode nccl, allowing unauthenticated attackers to exhaust KV-transfer worker memory. Attackers can call the exposed_set_value method to store unbounded key-value pairs without size limits, causing the worker process to crash and triggering node failure.
Executive ViewAI
LightLLM deployments up to version 1.2.0 that have been configured with the non-default NCCL prefill-decode transport mode are exposed to an unauthenticated remote attacker who can exhaust worker memory and take down serving nodes. The exposure is limited to availability - no customer data can be read or altered - but an outage would directly interrupt any inference-backed product or internal AI service running on the affected cluster, and could cascade if unhealthy nodes are not automatically replaced. Leadership should care because recovery depends on restarting or draining affected nodes and, at present, no vendor fix is available; risk is concentrated in teams that opted into the NCCL transport rather than in every LightLLM installation.
No vendor-released patch identified at time of analysis. Interim mitigations: remove or avoid the non-default --pd_trans_mode nccl flag (default transport mode is unaffected); restrict network access to the NCCL prefill-decode control channel to trusted hosts only via segmentation and firewall rules; run workers under an auto-restart supervisor with node health monitoring and capacity headroom; monitor worker memory consumption for anomalous growth and alert on unexpected process termination.
Threat Intelligence Overview
How this decision was reached
The engine reads each signal below and applies a fixed list of rules from top to bottom. The first rule that matches sets the recommended action – like a statement, every line is shown.
What every action level means
SSVC AssessmentENISA
Priority Score
Attack Technique
Denial of Service attacks render applications or systems unavailable by overwhelming resources or triggering failure conditions.
Unauthenticated denial of service in LightLLM through 1.2.0 lets any remote attacker who can reach the /generate endpoin
Unauthenticated remote code execution affects ModelTC's LightLLM LLM inference framework in all versions through 1.2.0,
Unauthenticated attackers who can reach the PD Master's /pd_register WebSocket endpoint can register arbitrary nodes in
Unauthenticated remote code execution affects LightLLM through 1.2.0 when it is deployed in visual_only (multimodal/visi
Unauthenticated remote code execution affects ModelTC LightLLM through version 1.2.0 when it is run in a multimodal depl
Remote code execution in LightLLM through 1.2.0 allows unauthenticated attackers to run arbitrary code on the host when
Remote code execution affects LightLLM deployments through version 1.2.0 that run the KV-transfer worker with the non-de
Unauthenticated HTTP requests to LightLLM through 1.2.0 can halt and wedge inference workers on any deployment started w
LightLLM through version 1.2.0 accepts image_url and audio_url parameters on its multimodal inference endpoints without
EUVD-2026-89483
GHSA-mqxw-j6vh-r94f
NVD
CVE.org
EPSS
Vulnrichment
Exploit-DB