Hub Host Security Guidance¶
This page is customer-facing security guidance for organizations deploying a QIE Hub. The Hub concentrates authority over every registered site, and that calls for explicit acknowledgement and for host hardening beyond a normal QIE engine deployment.
Treat this guidance as the minimum baseline. Defense in depth against sophisticated adversaries requires controls beyond what is documented here.
The Inherent Risk¶
A compromised Hub can reach every connected site's console through the tunnel, and it sees in cleartext everything that passes through it. At a site that has turned on Hub sign-in, it can also act as anyone the Hub vouches for, within the limits that site set. What a site has turned on decides how far a compromise reaches there.
What a Compromised Hub Can Reach states that reach precisely, site by site, with what still holds and what does not help. Read it before you deploy, and read it again before any site turns on Use Hub.
Treat the Hub as a Tier-1 Administrative Host¶
In your operational model, the Hub belongs in the same trust tier as:
- Active Directory domain controllers or equivalent identity systems.
- Privileged access management (PAM) bastions.
- Backup orchestration hosts.
It does not belong in the same tier as application servers or general workload hosts. Specifically:
- Host it on dedicated infrastructure, not shared with general application workloads.
- Restrict administrative access to the host (SSH, Remote Desktop, the vSphere console, etc.) to the operators who already hold tier-1 administrative access elsewhere in your environment. Do not extend it to QIE channel operators, hospital site administrators or general developers.
- Patch the Hub host's operating system, Java runtime and QIE on the same cadence as your other tier-1 hosts. Public vulnerabilities in any of them have a credible path to a Hub compromise.
Network Segmentation¶
The Hub exposes ports with different threat surfaces. Firewall each one accordingly:
| Port | Reachable from | Why |
|---|---|---|
| Dashboard (443) | Operator workstations or VPN | Authenticated human access only |
| Tunnel listener (8443) | The address ranges of registered QIE sites | Mutual TLS only; opens nothing to unauthenticated traffic |
| MCP Port, when the Hub's MCP server is on | The hosts that run your AI clients | HTTPS only; each client signs in through the Hub (see The MCP Server on a Hub) |
Specifically:
- Do not expose the tunnel listener port to the public internet unless you have no other option. If your sites reach the Hub over an MPLS, SD-WAN or customer VPN underlay, restrict the tunnel port to those source ranges at the network edge. Mutual TLS still blocks an unauthenticated attacker on that port, but every layer counts.
- Do not expose the dashboard port to the public internet in most deployments. Place it behind a VPN or a zero-trust proxy. The dashboard accepts password-style credentials, so a public dashboard invites credential stuffing and brute force.
- Filter egress. The Hub process should reach no more of the network than it needs. The only public internet destination QIE requires is the daily license check-in to
https://www.qvera.com. Everything else the Hub talks to is infrastructure you control: the database, and where configured, your identity provider (OIDC or LDAP), an SMTP relay for alerts, and log shipping.
Operator Authentication¶
The Hub uses QIE's own authentication (local users, LDAP or OIDC). For a Hub deployment:
- Prefer single sign-on with MFA through your enterprise identity provider (OIDC). Local passwords on a tier-1 host are a liability, because every operator's password becomes a target.
- Enforce account hygiene in your identity provider: MFA enrollment, password rotation, and disabling accounts at offboarding. The Hub adds no MFA of its own.
- Know that any Hub sign-in is enough for the Hub to vouch for a person. The Hub signs statements for a person however they signed in to the Hub. Statements for jobs and console sign-ins record the sign-in method, so the site's record shows it. One phished Hub password therefore reaches every site where that person's role holds a grant and the site accepts Hub sign-ins, within that site's limits.
- Limit the operator population and review it quarterly. A Hub user sees every site where their role holds the Console grant.
- Scope grants. Console opens a site's console through the Hub and signs the person in there; Jobs lets the person run jobs at the site; MCP lets an AI assistant work there as the person. Hub administrators hold all three only through their role, like anyone else. An administrator manages roles, though, so can give their own role any grant, and that change is written to the audit log; keep the administrator role small. Give grants for specific tags rather than All sites where you can, because a grant for All sites reaches each new site as soon as it enrolls. See Site Access Grants and Tags.
- Set the job approval threshold so that larger jobs need a second person (see Job Approval Threshold and Statement Retention). One Hub administrator can change the threshold alone, so approval guards against misuse of an operator account and against mistakes, not against a compromised administrator account.
Audit Log¶
The audit log is your primary forensic resource. Treat it as evidence:
- Ship
hub-audit.logoff the host continuously. See Shipping Off-Host. An attacker who reaches the Hub process can also reach its files; a copy in a SIEM you control is the integrity anchor. - Watch for
TUNNEL_TAKEOVER_DETECTEDstorms. Rapid repeated takeovers for one site are the main detection signal for a copied client certificate. See Tunnel Takeover Detection. - Alert on
BUNDLE_REUSE_REJECTED. A bundle submitted a second time means either an attacker who intercepted the bundle got there before the site, or an operator tried to reuse a consumed bundle. Either case warrants attention. - Alert on
BUNDLE_REVOKED_REJECTED. Something holding a bundle an operator deliberately canceled is trying to enroll. Treat it as an attempted enrollment by an unintended holder until you have established otherwise. - Alert on repeated
HUB_LOGIN_FAILUREfrom the same source address. - Review
GRANT_ADDEDandSITE_TAG_ADDEDevents. Each one widens where someone can reach or act through the Hub. - Review
JOB_SETTINGS_RESTOREDevents and the Hub Configuration revision history for changes to the approval threshold. - Compare
STATEMENT_SIGNEDlines inhub-audit.logwith what sites recorded. Each line is one statement the Hub signed. - Retain audit logs long enough for your regulatory obligations (HIPAA, state privacy law, contractual commitments). The file on the host rolls at 50 MB × 20 = 1 GB and cannot serve as an archive.
Site administrators should also watch their own sites' alerts. A site raises one when it refuses a statement that looks forged, replayed or misdirected.
Key Custody¶
The internal CA's private key and the statement-signing key are the most sensitive artifacts on the Hub. Treat anyone who can read the Hub database, or any backup or configuration export of it, as holding the CA key. See Hub Key Custody for what protects each key at rest and how to rotate the CA.
Path-Based Routing (a Security Trade-Off)¶
Subdomain-per-site routing is strongly recommended because it gives each site its own browser origin. The optional path-based mode (-Dqie.hubProxyAllowPathBased=true) serves every site under the Hub's single origin instead, which removes that isolation:
- Every site shares the Hub host's browser origin.
- Cookies and JavaScript can read across sites the same operator has open in different tabs.
- A compromised site can run arbitrary JavaScript against another site's open console.
Path-based routing is provided for deployments that cannot provision per-site wildcard DNS and a wildcard TLS certificate. If you enable it, you accept these risks in exchange for that simplicity. Make it a deliberate, documented decision, and prefer wildcard DNS and a wildcard certificate with subdomain routing wherever you can.
During your deployment review, confirm whether path-based routing is enabled on each Hub instance, so the setting is intentional rather than accidental:
Recommended Additional Controls¶
These are not required to run a Hub, but they materially improve resilience to specific threats:
- Disk encryption on the Hub host. Standard for tier-1 hosts.
- Process-level egress filtering (cgroups, systemd network policy, a host firewall with per-binary rules). The only outbound internet call QIE must make is the daily license check-in to
https://www.qvera.com, which combines licensing, update notification and usage reporting. If it is missed for long enough, the Hub warns and then locks down (see Hub Lockdowns). Allow that one destination and deny the rest of the public internet. The optional automatic download of a new build also goes toqvera.com; block it and apply updates manually if you prefer. - Monitoring of sensitive audit events in close to real time. A rule that pages on-call for certificate revocation, grants added to roles, tags added to sites and bundle generation can catch an unauthorized operator action as it happens.
- Configuration drift detection on the
hub_listener_configrow, thehub_statement_signing_keyrow and the SSL certificate and key tables. Unexpected changes to the server certificate, the internal CA, the external URLs, the Hub id or the approval threshold deserve investigation. - Periodic review of the deployment. Quarterly: the operator population, role grants, tags and the revocation list. Annually: a full security review against your threat model.
AI Assistants Through the Hub¶
A Hub runs its own MCP server. An AI client signs in once, at the Hub, and the person approves it there. The Hub proxies a site's console, never the site's own MCP port, and a site refuses MCP client sign-in for any request the Hub proxied.
Each tool call an AI client makes at a site travels as its own statement, and runs at the site as the person's Hub user, under the site's own MCP rules, limits and audit log. A site takes these calls only when Use Hub and Accept MCP calls through the Hub are both on. Through the Hub, the user and role tools are refused, and so is a zone save that sends users or roles. See The MCP Server on a Hub, MCP Calls Through the Hub and MCP Server Security Guidance.
What Is Not in Scope for This Guidance¶
- Insiders with legitimate Hub operator access who abuse it. The audit log records what they did; deterrence and personnel controls belong to your organization.
- Compromise of an operator's workstation. If an operator's browser is compromised, every Hub session open in it is compromised too. Standard endpoint hygiene applies.
- Compromise of a site's own host. A compromised site can report false status to the Hub, or take PHI out through actions an operator is permitted to perform there. Site host hygiene is outside the Hub's scope.
Reporting a Vulnerability¶
Report security issues in the QIE Hub, or any QIE component, privately to Qvera through the established responsible disclosure channel. See your support contract for the dedicated security contact. Do not file public issues for unpatched vulnerabilities.