Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessReach exposed PD Master /pd_register WebSocket
-
DeliverySend crafted JSON node-registration payload
-
ExploitBypass missing auth and peer-address check
-
ExecutionRegister rogue prefill/decode node
-
ImpactExfiltrate routed user prompts / evict real nodes / SSRF to internal hosts
Vulnerability Assessment
| Weakness |
CWE
OWASP Top 10:2025
A07:2025 - Authentication Failures
ASVS 5.0
V6 - Authentication Authentication required for critical functions
MITRE ATT&CK
1 technique across 1 tactic: Initial Access The technique list with ATT&CK links and the kill-chain view is part of the Basic plan and up. |
| Exploitation Conditions | Requires that the target run LightLLM (≤1.2.0) in PD (prefill-decode) disaggregated serving mode with the PD Master's /pd_register WebSocket endpoint network-reachable to the attacker; per CWE-306 and PR:N no credentials are needed and no user interaction is required, and the flaw is that peer-address validation is absent on node registration. Limiting factors: this endpoint exists only in the multi-node PD Master topology (not the default single-node deployment), so single-node or network-isolated deployments where the endpoint is firewalled off are not exposed - the practical constraint is endpoint reachability, not attacker privilege. |
| Impact | This is a genuine high-priority issue, not a high-CVSS-but-low-real-risk case. The vendor CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H, score 9.3) describes a remotely reachable, unauthenticated, low-complexity flaw with high confidentiality, integrity, and availability impact, and this is consistent with the CWE-306 (Missing Authentication) classification and the described capabilities (prompt disclosure, node-replacement DoS, and SSRF to internal addresses). Unlike many kernel bugs where high vendor scores mask crash-only impact, here the impacts are real and multi-faceted: attacker-driven prompt exfiltration is a true confidentiality breach and SSRF into internal networks is a real integrity/pivot primitive. Signals are partially missing: there is no EPSS score, no CISA KEV listing, and no public POC in the provided data - so exploitation is currently 'no public exploit identified at time of analysis' despite the high score, suggesting the score reflects impact potential rather than confirmed in-the-wild activity. The main real-world discount is deployment-conditional: the vulnerable /pd_register endpoint lives in the PD (prefill-decode disaggregated) Master serving mode, which is a specific multi-node topology rather than LightLLM's default single-node deployment, so the population of exposed, vulnerable instances is smaller than the raw AV:N/PR:N vector implies. Verify exposure by checking whether the PD Master WebSocket endpoint is network-reachable. |
| Exploit ScenarioDerived | An attacker who can reach a LightLLM PD Master's /pd_register WebSocket endpoint sends crafted JSON to register a rogue prefill/decode node without any authentication or peer-address validation. Once registered, user prompts get routed to the attacker's socket (disclosing sensitive prompt content), the attacker can evict legitimate nodes to cause denial of service, or coerce the PD Master into issuing requests to internal network addresses (SSRF). No public exploit code is identified in the provided data, but the VulnCheck advisory and GitHub issue #1576 document the flaw and pinpoint the vulnerable code paths. |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Sep 18, 2026
Unauthenticated attackers who can reach the PD Master's /pd_register WebSocket endpoint can register arbitrary nodes in LightLLM through 1.2.0, causing full user prompts routed to the attacker's socket to be disclosed, disrupting service by displacing legitimate nodes, and forcing the PD Master to issue requests toward internal network addresses. The root cause is CWE-306 (missing authentication for a critical function): the registration handler accepts crafted JSON without validating the peer address, and the CVSS 4.0 score of 9.3 with PR:N/UI:N reflects high confidentiality, integrity, and availability impact requiring no credentials or user interaction. No public exploit code or confirmed active exploitation was identified at time of analysis, and practical exposure is limited to multi-node prefill-decode (PD) disaggregated deployments where the endpoint is network-reachable; single-node or properly firewalled deployments are not exposed.
RemediationAI
No vendor-released patch was identified at time of analysis, so the primary action is to track the upstream issue https://github.com/ModelTC/LightLLM/issues/1576 and the repository https://github.com/ModelTC/LightLLM for a fixed release and upgrade beyond 1.2.0 as soon as one is published; until then, treat the PD Master as an untrusted-exposure control plane. The most effective compensating control is network isolation: restrict the /pd_register WebSocket listener with firewall rules or security groups so only known prefill/decode node addresses can connect, ideally binding it to an internal interface rather than a public one - the trade-off is that dynamic autoscaling or node replacement outside the allowed address set will fail registration, so allowlists must be maintained alongside cluster changes. Where topology flexibility is required, terminate the WebSocket behind a reverse proxy that enforces mutual TLS or an authentication token and validates the client peer address before forwarding to the PD Master, accepting the added operational complexity of certificate/token distribution. Avoid exposing the PD Master directly to any untrusted network or the public internet, and monitor PD Master logs for unexpected node registrations or registrations from unfamiliar addresses, which would indicate exploitation attempts; these detection measures do not remove the vulnerability and should be paired with segmentation until a patch is available.
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's prefill-decode (PD) disaggregation architecture separates prefill and decode work across multiple nodes, with a central PD Master coordinating membership; nodes join by connecting to the PD Master's /pd_register WebSocket endpoint. In version 1.2.0 and earlier, that registration path performs no authentication and, critically, no peer-address validation - the handler in lightllm/server/httpserver_for_pd_master/manager.py (registration logic around lines 691-749) trusts attacker-supplied JSON describing a node, and the endpoint wiring appears in lightllm/server/api_http_pd.py (lines 26-36) and lightllm/server/api_http.py (lines 198-205). This is the classic CWE-306 pattern: a security-critical control-plane function (node registration and routing table modification) is exposed without any authentication check, so anyone able to open the WebSocket can insert themselves into the serving topology. Because the PD Master subsequently routes inference traffic to registered nodes, a rogue registration yields prompt disclosure and lets the attacker replace legitimate nodes (availability impact) while also causing the PD Master to make outbound requests to attacker-chosen internal addresses, effectively turning the control plane into an SSRF vector. The affected CPE cpe:2.3:a:modeltc:lightllm:*:*:*:*:*:*:*:* confirms all releases up to and including 1.2.0 are in scope; the independent assessment (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H) matches the provided CVSS 4.0 vector in impact and prerequisites, with the key precondition being network reachability to the endpoint rather than any attacker privilege.
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:H/VI:H/VA:H/SC:N/SI:N/SA:N
Network-reachable unauthenticated WebSocket (AV:N/PR:N/UI:N/AC:L); full prompt disclosure (C:H), rogue node registration and SSRF (I:H), and node-replacement DoS (A:H); scope unchanged as impact is on the LightLLM service.
Primary rating from Vendor (VulnCheck).
CVSS VectorVendor: VulnCheck
DescriptionCVE.org
LightLLM through 1.2.0 contains an authentication bypass vulnerability in the /pd_register WebSocket endpoint that allows unauthenticated attackers to register arbitrary nodes by supplying crafted JSON without peer address validation. Attackers can disclose full user prompts routed to their socket, trigger denial of service by replacing legitimate nodes, or make the PD Master issue requests to internal network addresses.
Executive ViewAI
LightLLM deployments that run multi-node prefill-decode (PD) disaggregation are exposed to an unauthenticated attacker who can silently hijack the node registration channel, causing customer and internal prompts to be routed to attacker-controlled infrastructure and disrupting inference availability. Because the flaw requires no credentials and no user interaction, any organization running network-reachable PD clusters risks both sensitive data disclosure and service outage. Leadership should care because the impacted asset is a core AI inference path that typically processes proprietary or regulated data, and there is no vendor fix to fall back on.
No vendor-released patch identified at time of analysis. Interim mitigations: network-isolate the /pd_register WebSocket endpoint so only trusted PD node addresses can reach it; enforce mTLS and peer-address validation on node registration; segment PD Master and workers onto a dedicated management network with restricted egress; monitor and alert on unauthorized node registrations and anomalous outbound requests to internal addresses. Note: unauthenticated (CVSS PR:N/UI:N); no public exploit code or confirmed active exploitation identified at time of analysis; exposure limited to network-reachable multi-node PD disaggregated deployments.
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
Authentication bypass attacks exploit flaws in the verification mechanisms that control access to systems and applications.
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 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
Memory exhaustion in LightLLM through 1.2.0 lets unauthenticated remote attackers crash KV-transfer worker processes whe
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-2026-83283
GHSA-644q-7756-5g89
NVD
CVE.org
EPSS
Vulnrichment
Exploit-DB