Scope
QHx establishes workload identities from configured node and workload evidence, secures proxy-mediated communication, and can preserve selected interactions through supported Notary middleware. Resource admission, credential issuance, connection authentication and request recording are separate control points.
The deployment must also provide host hardening, protected root keys, software-supply-chain controls, privileged-access governance and application security. QHx does not manage human identity governance or establish approval to transfer data between classification domains. Those transfers require the applicable authorized cross-domain mechanism.
The architecture overview identifies the components. The guides on identity, communication, evidence and operations describe their control points and configuration decisions.
Adversaries
The threat model considers malicious workloads attempting credential abuse or lateral movement; compromised hosts or hypervisors; operators abusing legitimate privileges; network attackers intercepting or modifying traffic; compromised software introduced through a supply chain; and cross-domain trust confusion. It also considers adversaries retaining encrypted traffic for future cryptanalysis.
These adversaries have different reach. A network attacker does not have the same capabilities as a host administrator. A control that protects the connection cannot be credited with protecting keys already exposed by a compromised endpoint.
Controls and dependencies
| Threat | QHx control | Remaining dependency |
|---|---|---|
| Workload impersonation | Node and workload evidence gate credential issuance. Short-lived SVIDs limit credential validity. | Protect the host, private keys, registration rules and issuing authority. Extracted key material can still enable impersonation. |
| Unauthorized communication | Configured proxy paths authenticate peers and apply identity restrictions. Admission labels can drive generated NetworkPolicies. | Protect policy authority and enforce the network rules. Test permitted paths, denied paths and routes that bypass the proxy. |
| Traffic interception | Mutually authenticated, encrypted tunnels protect configured connections. | Application vulnerabilities, compromised endpoints and malicious requests from authorized peers remain outside transport protection. |
| Unapproved resource changes | Signed resources verify an accepted signing authority and version before application. | Protect signing keys and trusted version state. An authorized signature does not establish that a change is safe. |
| Disputed request records | Configured Notary middleware can sign selected captured fields. The HTTP capture path has endpoint and response constraints. | Retain the record and required trust material. Signer, host, time and capture assumptions determine what the evidence establishes. |
| Trust confusion across domains | Explicit SPIFFE federation defines which foreign issuing authorities can be recognized. | Authenticate bundle distribution, manage freshness, and apply separate service and data authorization. |
| Future decryption of recorded traffic | ML-KEM and hybrid key-establishment profiles provide post-quantum options for participating connections. | Both peers must use compatible profiles. Classical recordings retain their original exposure; endpoints and keys still require protection. |
Separate components do not automatically create separation of administrative authority. Operators must enforce privileges over the Manager, PKI, Notary, signing keys and policy state. An administrator who can alter all of these controls can undermine their intended boundaries.
Residual risks
Host and authority compromise
A compromised host can impersonate workloads running on it. A certificate and its private key may remain usable while accepted, even after the original process stops. Short validity periods do not erase extracted material or guarantee that an established session ends at certificate expiry.
Compromise of an issuing root or authorized signing key can create evidence that otherwise passes verification. Protecting root keys, administrative access and recovery state remains a deployment responsibility.
Limits of attestation and recorded evidence
TPM measurements apply where that attestation path is enabled. They strengthen particular issuance checks; they do not prove continuous host integrity or that all executing code is benign. Other attestation methods provide different evidence and must be assessed against the intended threat.
Attestation and notarization do not replace trusted builds, artifact review or application testing. A signed malicious artifact can still carry an authorized signature. A signed request record supports verification of captured material under its signer; it does not independently establish truthful execution, the correctness of an AI result, or an unstated model version.
Policy, freshness and released data
Overbroad policy, incorrect selectors or permissive federation can authorize the wrong workload. During disconnection, cached policy and trust bundles can become stale; a signature cannot communicate a change that has not reached the deployment. Credential expiry, local renewal and recovery behavior need explicit operational decisions.
CABE key access and data-release policy are separate from peer authentication. Changes in permission or expiry do not erase plaintext or keys already released to a recipient. An evaluation should test those boundaries alongside allowed exchanges and failure recovery.
Loss of reach-back
If the connection to a remote operating site fails, a local application exchange can continue only while its local network, proxies, accepted credentials, policy and trust material remain usable. Renewing an identity additionally needs an available issuer and its required attestation services. A cached certificate or a verifiable receipt cannot supply that service.
Exercise the intended outage with the actual credential lifetimes and local dependencies. Check new connections and established sessions separately, then test trust refresh and recovery after reconnection. The operations guide describes the preparation.
Object keys and federated authorities
A QHx proxy’s connection policy does not authorize every application action or disclose an object’s key. In a CABE integration, the Key Service authenticates the client and applies its own Policy to key requests. SPIFFE identity federation recognizes an issuer; CABE Key Service federation can share key material with a receiving service. Those are different grants of trust.
A CABE Domain is not inherently a classification domain or an approved security boundary. Encrypting an object for carriage through a relay does not establish approval to place the information on that transport. The publishing organization must apply its approved sharing arrangements.
An adapter that transforms plaintext is an authorized processor, not merely an opaque relay. Its identity identifies that processor; it does not repair missing authenticated origin at the input. Shared-key authenticated encryption also does not uniquely identify a producer among parties holding the same key.
For the illustrated object path, a broker without the key cannot read the Payload, but it can still delay, drop or replay Envelopes and observe traffic. Key Services and participants that hold usable keys remain trusted for those keys. See CABE’s security model.