CVE-2026-102895: GPUStack Localhost Authentication Bypass to Remote Code Execution and Model Tampering
A critical vulnerability (CVE-2026-102895) in GPUStack before v2.2.0 permits remote unauthenticated adversaries to acquire full administrative control over distributed GPU AI clusters. GPUStack is widely deployed across enterprise private clouds to orchestrate local LLM serving (vLLM, Ollama, llama.cpp) and multi-agent AI execution pipelines. A flaw in its loopback authentication handler (gpustack/api/auth.py) unconditionally trusts requests perceived as originating from localhost (127.0.0.1). Attackers forge this identity using unverified 'X-GPUStack-Real-IP' headers or by taking advantage of standard reverse-proxy forwarding topologies. With administrative access, attackers can modify or replace model weights, exfiltrate private prompt contexts and embeddings, generate persistent administrative API tokens, and achieve remote code execution via worker provisioning sinks. CISA added CVE-2026-102895 to the Known Exploited Vulnerabilities (KEV) catalog on November 9, 2026. Hermes rates this threat at HTS 96 (EXTREME) and HASS 94 (EXTREME).
HERMES THREAT SCORE & AI INFRASTRUCTURE RESILIENCE EVALUATION
Target:GPUStack Model Orchestration & Worker Management Plane CVSS v3.1 rates CVE-2026-102895 at 9.8 (Critical). Hermes assigns an extreme HTS score of 96 and an Agentic Security Score (HASS) of 94. Beyond host compromise, GPUStack is the operational brain for enterprise LLM agents and retrieval-augmented generation (RAG) pipelines. An adversary controlling GPUStack silently poisons model weights, injects covert backdoors into system prompts, and intercepts proprietary corporate queries at the inference boundary with zero friction.
1. Executive Summary & Context
Section titled โ1. Executive Summary & ContextโGPUStack is an open-source model serving and cluster management platform designed to orchestrate GPUs and NPUs across distributed nodes for running large language models (LLMs) and multi-agent AI frameworks. It exposes unified OpenAI-compatible endpoints and automates model deployment, quantization, and resource scheduling.
On November 6, 2026, security researchers identified a catastrophic authentication bypass (CVE-2026-102895) in GPUStack before version 2.2.0. By exploiting unvalidated loopback trust logic in gpustack/api/auth.py, external remote attackers can present arbitrary requests as originating from localhost, causing the application to automatically authenticate them as the built-in system administrator without credentials. On November 9, 2026, CISA formally added CVE-2026-102895 to the Known Exploited Vulnerabilities (KEV) catalog with evidence of threat actors deploying reverse shells and exfiltrating proprietary AI models in the wild.
2. Root Cause Analysis & Technical Mechanics (CWE-287 / CWE-290 / CWE-346)
Section titled โ2. Root Cause Analysis & Technical Mechanics (CWE-287 / CWE-290 / CWE-346)โThe root cause resides in GPUStackโs get_current_user() dependency handler located in gpustack/api/auth.py.
Flawed Loopback Auto-Admin Bypass
Section titled โFlawed Loopback Auto-Admin BypassโIn default installations, GPUStack was designed with a convenience feature allowing administrators accessing the dashboard locally to skip authentication (force_auth_localhost=False by default). The authentication logic verified the client address as follows:
# Vulnerable snippet in gpustack/api/auth.pyasync def get_current_user( request: Request, token: Optional[str] = Depends(oauth2_scheme), session: AsyncSession = Depends(get_session),) -> User: client_ip = request.headers.get("X-GPUStack-Real-IP") or request.client.host
# Insecure convenience bypass: if not system_config.force_auth_localhost and client_ip in ("127.0.0.1", "::1"): admin_user = await get_system_admin(session) if admin_user: return admin_user
# Standard bearer token validation return await authenticate_with_token(token, session)Two critical flaws converge to make this trivially exploitable remotely:
- Unsanitized Header Trust: The server reads
X-GPUStack-Real-IPdirectly from client request headers without validating whether the connection arrived through a trusted reverse proxy. Any remote attacker supplyingX-GPUStack-Real-IP: 127.0.0.1causesclient_ipto resolve to loopback. - Reverse Proxy Masking: In deployments where GPUStack is fronted by Nginx, Traefik, or an AWS Application Load Balancer without explicit
proxy_set_headeroverrides, the incoming socket peer IP (request.client.host) itself is127.0.0.1, inadvertently triggering the bypass for all external callers.
Privilege Escalation to Host RCE
Section titled โPrivilege Escalation to Host RCEโOnce granted the admin_user identity, the attacker accesses all administrative endpoints:
POST /v1/api-keys: Creates permanent API keys with full administrative scope.POST /v1/models: Deploys trojaned models or alters inference parameters.POST /v1/workers: Provisions new worker nodes or triggers worker registration scripts, executing arbitrary operating system commands under the permissions of the GPUStack service user.
3. Threat Architecture & Attack Vectors
Section titled โ3. Threat Architecture & Attack VectorsโsequenceDiagram autonumber actor Attacker as Threat Actor participant Proxy as Reverse Proxy / Ingress participant GPUStack as GPUStack Server (auth.py) participant Cluster as Distributed GPU Workers participant Storage as Model Weight Store
Attacker->>Proxy: POST /v1/api-keys<br/>Header: X-GPUStack-Real-IP: 127.0.0.1 Proxy->>GPUStack: Forward request (Peer: 127.0.0.1) GPUStack->>GPUStack: get_current_user() checks client_ip<br/>client_ip == "127.0.0.1" (Bypass Active!) GPUStack-->>Attacker: 201 Created (Admin API Key generated)
Attacker->>GPUStack: POST /v1/workers (Deploy rogue worker script) GPUStack->>Cluster: Execute worker setup command Cluster-->>Attacker: Interactive Reverse Shell (RCE achieved)
Attacker->>GPUStack: GET /v1/models & POST /v1/models GPUStack->>Storage: Pull/Tamper model weights & prompt templates Storage-->>Attacker: Exfiltrated LLM Weights & Tenant PromptsExploit Verification Flow
Section titled โExploit Verification FlowโAn attacker verifies vulnerability status with a single curl command:
curl -s -X GET "http://target-cluster:10150/v1/users/me" \ -H "X-GPUStack-Real-IP: 127.0.0.1" \ -H "Accept: application/json"If vulnerable, the server responds with HTTP 200 and the JSON record of the built-in admin account ({"id": 1, "username": "admin", "is_admin": true}).
4. Forensic Triage & Detection Engineering
Section titled โ4. Forensic Triage & Detection Engineeringโtitle: GPUStack Localhost Authentication Bypass Attemptid: 5a4f7812-9c10-4e2b-b8f2-39c87890a102status: experimentaldescription: Detects HTTP requests targeting GPUStack with forged localhost headers or unauthenticated administrative calls.author: Hermes Codex Cyber Defense Teamdate: 2026-11-10references: - https://nvd.nist.gov/vuln/detail/CVE-2026-102895 - https://www.cisa.gov/known-exploited-vulnerabilities-cataloglogsource: category: webserverdetection: selection_headers: c-header|contains: - 'X-GPUStack-Real-IP: 127.0.0.1' - 'X-GPUStack-Real-IP: ::1' - 'X-GPUStack-Real-IP: localhost' selection_endpoints: cs-method: 'POST' cs-uri-stem|startswith: - '/v1/api-keys' - '/v1/users' - '/v1/workers' - '/v1/models' condition: selection_headers or selection_endpointsfields: - c-ip - cs-method - cs-uri-stem - c-headerfalsepositives: - Internal testing scripts originating directly from the host console without proper proxy stripping.level: criticaltags: - attack.initial_access - attack.t1190 - attack.persistence - attack.t1078.001// Detect anomalous GPUStack administrative calls with loopback header spoofingW3CIISLog| where csUriStem startswith "/v1/"| where csHeader contains "X-GPUStack-Real-IP" or csHeader contains "127.0.0.1"| extend ClientIP = cIP, RequestPath = csUriStem, UserAgent = csUserAgent| summarize Count = count(), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by ClientIP, RequestPath, UserAgent| where Count > 0| order by LastSeen desc#!/bin/bash# GPUStack Forensic Audit Script for CVE-2026-102895echo "[*] Auditing GPUStack log files for localhost spoofing indicators..."
LOG_PATH="/var/log/gpustack"if [ -d "$LOG_PATH" ]; then grep -E -i "X-GPUStack-Real-IP.*(127\.0\.0\.1|::1)" "$LOG_PATH"/*.log grep -E -i "User 'admin' authenticated via localhost bypass" "$LOG_PATH"/*.log grep -E -i "POST /v1/(api-keys|workers|models)" "$LOG_PATH"/*.log | grep -v "200"else echo "[-] GPUStack log directory not found at $LOG_PATH. Checking journald..." journalctl -u gpustack --since "7 days ago" | grep -E "127\.0\.0\.1|localhost|bypass"fi5. Remediation & Hardening Strategy
Section titled โ5. Remediation & Hardening Strategyโ-
Immediate Upgrade to GPUStack 2.2.0 or 2.2.2: Deploy the official patch released by the GPUStack development team. Version 2.2.0 strips the insecure auto-admin path and requires API token validation across all interfaces.
Terminal window pip install --upgrade gpustack>=2.2.0systemctl restart gpustack -
Workaround โ Force Localhost Authentication: If immediate patching is impossible, enforce strict authentication for loopback traffic in GPUStack configuration:
Terminal window # Add to /etc/gpustack/config.yaml or launch flagsforce_auth_localhost: true -
Reverse Proxy Header Stripping: Configure all border reverse proxies (Nginx, Traefik, Envoy, AWS ALB) to strip incoming client headers:
# Nginx configurationproxy_set_header X-GPUStack-Real-IP $remote_addr;proxy_hide_header X-GPUStack-Real-IP; -
Credential and Model Audit: Rotate all GPUStack administrative tokens, audit registered worker nodes in
gpustack.db, and perform cryptographic checksum verification on stored model weights (.gguf,.safetensors).