CVE-2026-129428: Pulse Infrastructure AI Assistant Approval API Authorization Bypass
On November 15, 2026, VulnCheck published weaponized exploit research disclosing CVE-2026-129428, an authorization bypass flaw in Pulse (rcourtman/pulse versions 5.0.17 through 6.7.x). Pulse is a unified infrastructure monitoring and orchestration platform managing Proxmox VE hypervisors, Docker runtimes, Kubernetes clusters, and TrueNAS storage nodes. The platform incorporates an autonomous AI assistant capable of suggesting and automating operational tasks. Sensitive administrative actionsโsuch as stopping core virtual machines, deleting storage pools, or deploying network reconfiguration scriptsโare queued in a Human-in-the-Loop (HITL) approval registry. However, the approval API endpoints (/api/ai/approvals) fail to enforce role-based access controls. Any authenticated user holding basic viewer-only privileges can query pending high-impact action proposals, inspect proposed parameters, and issue unauthorized approvals. This turns the autonomous AI assistant into an unwitting, high-privilege proxy for cluster sabotage. Immediate upgrade to Pulse version 6.8.0 or later is required.
HERMES THREAT SCORE & ENTERPRISE INFRASTRUCTURE RISK
Target:Pulse Platform AI Assistant Approval Endpoints and Proxmox/Kubernetes Execution Dispatcher CVSS v3.1 rates this flaw at 7.5 (High) or 5.4 (Medium) under classical web vulnerability models, treating it as an ordinary role check omission. Hermes evaluates the threat through agentic risk modeling: the AI assistant operates with hypervisor and cluster root authority. When an unprivileged viewer bypasses Human-in-the-Loop (HITL) approval gates, the AI agent becomes a high-privileged execution proxy capable of dismantling virtualized datacenter infrastructure without triggering traditional privileged command audit alarms.
HASS AGENTIC SEVERITY & AUTONOMOUS ACTION DISPATCH ABUSE
Target:Autonomous AI Assistant Action Queue & HITL Authorization Gate In autonomous IT operations, Human-in-the-Loop (HITL) authorization represents the primary safety tripwire preventing autonomous agents from inflicting irreversible system damage. By eliminating the requirement for administrator credentials on the approval endpoint, the agent's broad toolsetโincluding Proxmox VM termination, ZFS volume destruction, and container orchestrationโis exposed directly to low-privilege actors, collapsing the core safety boundary of the agentic control plane.
1. Executive Summary & Operational Context
Section titled โ1. Executive Summary & Operational ContextโPulse (rcourtman/pulse) is a self-hosted real-time infrastructure monitoring and operational dashboard heavily utilized in private clouds, homelabs, and edge datacenter environments managing Proxmox VE, Docker, Kubernetes, and TrueNAS.
With the introduction of native AI assistant capabilities in versions 5.x and 6.x, Pulse integrates LLM-driven autonomous tooling to automate complex cluster maintenance tasks. Because the assistant connects directly to underlying hypervisor and container daemon APIs with root or cluster-admin API tokens, high-risk actionsโsuch as shutting down production virtual machines, pruning persistent volumes, or executing shell scriptsโare designed to require explicit approval through an interactive Human-in-the-Loop (HITL) workflow.
On November 15, 2026, VulnCheck disclosed CVE-2026-129428. The backend API routes handling AI approval management (/api/ai/approvals) fail to verify user permissions. An authenticated user possessing only read-only or viewer access can view pending operational proposals and unilaterally approve them. This enables low-privilege users or compromised low-level accounts to bypass administrative authorization and weaponize the AI assistant to perform catastrophic infrastructure mutations.
2. Technical Vulnerability Analysis & Root Cause
Section titled โ2. Technical Vulnerability Analysis & Root CauseโThe flaw exists within the approval controller routing definition (src/server/routes/api/ai/approvals.ts).
While standard administrative endpoints in Pulse enforce role-based middleware (requireAdminRole), the newly added /api/ai/approvals endpoints were mounted with generic authentication middleware (requireAuth) that only checks for a valid session token, omitting the role check:
// Vulnerable routing pattern in src/server/routes/api/ai/approvals.tsimport { Router } from "express";import { requireAuth } from "../../middleware/auth"; // Only verifies session presence!
const router = Router();
// Missing requireAdminRole middleware!router.get("/api/ai/approvals", requireAuth, async (req, res) => { // Returns all pending high-privilege action proposals, including shell commands and targets const pending = await db.aiApprovals.find({ status: "PENDING" }); return res.json({ proposals: pending });});
router.post("/api/ai/approvals/:id/decision", requireAuth, async (req, res) => { const { decision } = req.body; // "APPROVED" or "REJECTED" const { id } = req.params;
// Flaw: No role verification! A viewer-level user can approve! const proposal = await db.aiApprovals.findById(id); if (!proposal) return res.status(404).json({ error: "Not found" });
proposal.status = decision; proposal.reviewedBy = req.user.username; // Viewer recorded as approver! await proposal.save();
if (decision === "APPROVED") { // Triggers execution dispatcher with root Proxmox / Kubernetes credentials await actionDispatcher.execute(proposal.actionPayload); }
return res.json({ success: true, status: decision });});Because requireAuth validates only that the caller is logged in, any user with the viewer role can submit a POST /api/ai/approvals/<id>/decision payload with {"decision": "APPROVED"}. The platform then immediately triggers actionDispatcher.execute(), executing destructive operations across hypervisors with full cluster authority.
3. Attack Execution Flow & Weaponization Mechanics
Section titled โ3. Attack Execution Flow & Weaponization Mechanicsโflowchart TD A["Low-Privilege User / Compromised Account<br/>(Role: viewer)"] -->|"Step 1: Authenticates with Basic Credentials"| B["Pulse Web Dashboard API"] B -->|"Step 2: GET /api/ai/approvals<br/>Omits Role Check"| C["Discovers Pending AI Proposals<br/>(e.g., destroy_vm, prune_volumes)"] A -->|"Step 3: POST /api/ai/approvals/:id/decision<br/>Payload: {'decision': 'APPROVED'}"| D["AI Approval Controller"] D -->|"Step 4: Accepts Viewer Approval<br/>without requireAdminRole"| E["Action Execution Dispatcher"] E -->|"Step 5: Invokes Hypervisor / K8s APIs<br/>with Cluster-Admin Credentials"| F["Proxmox VE & Kubernetes Nodes"] F -->|"Step 6: Destructive Action Executed"| G["Data Loss & Outage across Virtual Infrastructure"]4. Forensic Signatures, Detection Queries & IOCs
Section titled โ4. Forensic Signatures, Detection Queries & IOCsโtitle: Pulse AI Assistant Unauthorized Approval Submissionid: b1294280-1294-4680-9129-129428000001status: experimentaldescription: Detects approval decisions submitted to Pulse AI Assistant endpoints by non-admin or viewer accountslogsource: category: webserver product: pulsedetection: selection_path: c-uri|contains: '/api/ai/approvals/' c-method: 'POST' selection_decision: cs-method: 'POST' cs-uri-stem|endswith: '/decision' condition: selection_path and selection_decisionlevel: high// Detect HTTP requests to Pulse AI approvals originating from non-admin user sessionsW3CIISLog| where csUriStem has "/api/ai/approvals" and csMethod =~ "POST"| where csUriStem endswith "/decision"| extend ApprovingUser = csUsername| project TimeGenerated, cIP, csMethod, csUriStem, ApprovingUser, scStatus# Grep for approval events triggered by non-admin accounts in Pulse logsgrep -iE "approval.*granted|reviewedBy" /var/log/pulse/app.log | grep -ivE "admin|root"
# Review recent Proxmox VM lifecycle actions triggered by API tokensjournalctl -u pve-cluster -u pvedaemon --since "2 days ago" | grep -iE "qm destroy|qm stop|pct destroy"
# Audit Pulse user roles in SQLite or MongoDB databasesqlite3 /var/lib/pulse/pulse.db "SELECT id, username, role, last_login FROM users;"Threat Intelligence & Indicators of Compromise (IOCs)
Section titled โThreat Intelligence & Indicators of Compromise (IOCs)โ| Type | Indicator | Context |
|---|---|---|
| IP Address | 141.98.11[.]88 | Automated crawler probing for unprotected Pulse dashboards |
| IP Address | 195.201.225[.]14 | Secondary scanner testing /api/ai/approvals endpoints |
| API Route | /api/ai/approvals/:id/decision | Target endpoint vulnerable to unprivileged approval |
| Audit Log Event | actionDispatcher.execute() | Execution log entry corresponding to viewer approval |
5. Comprehensive Remediation, Hardening & Mitigation Plan
Section titled โ5. Comprehensive Remediation, Hardening & Mitigation Planโ-
Upgrade to Pulse Version 6.8.0 or Later: Deploy Pulse version 6.8.0 immediately. The update attaches the
requireAdminRolemiddleware to all/api/ai/approvalsroutes, enforcing administrative privileges before any proposal can be inspected or decided. -
Revoke Unused Viewer Accounts: Audit active users in the Pulse dashboard. Disable guest registration and remove inactive accounts with access to monitoring endpoints.
-
Segment Agent Execution Privileges: Apply the principle of least privilege to the API tokens provided to Pulse for Proxmox VE, Docker, and Kubernetes. Restrict the AI assistantโs service account to read-only status commands, removing
VM.DestroyandStorage.Allocateprivileges unless strictly necessary. -
Review Historic Execution Audit Logs: Examine
/var/log/pulse/and hypervisor task logs (pve-cluster) to confirm whether any pending proposals were approved by unexpected users prior to upgrading.