Endpoint Policy

Understand endpoint and transport policy

An AI or node endpoint is a security boundary because the app may send documents, commands or identity material to it. EdgeORE's policy rejects public AI destinations for the owned-host path and validates resolved addresses. The client connects to a validated address rather than validating a hostname and then resolving it again without control. This reduces a class of destination changes, but it is not a complete ownership or network-security proof.

Validate all destinations

A hostname can resolve to multiple addresses. The policy needs to consider every resolved address, including IPv4 and IPv6 results. A mixed result containing a disallowed public address should not be accepted merely because another result is private. Redirects are another way a request could leave its approved destination; the node and RPC clients refuse them under the candidate's hardened transport behavior.

Loopback cleartext is allowed for a tightly scoped development path. That exception should not be described as all private-network traffic being encrypted. LAN usage requires the configured HTTPS behavior. Read the binary network-security configuration and runtime results when qualifying a build; source intent alone is not a device TLS observation. A hosted emulator's connectivity may also differ from an owned phone on the same LAN as the host.

Test policy and reachability separately

A public endpoint should be refused before the app sends sensitive request content. An allowed loopback endpoint can still be unreachable if the reverse port is absent or the model server is down. An allowed private hostname can still fail certificate validation. Those cases need distinct messages because changing the wrong setting can weaken the security boundary without fixing the connection.

When diagnosing, record the policy decision and the transport error without logging document content, authorization tokens or private keys. A redacted endpoint may still be useful if its classification and port are preserved. Avoid publishing personal internal hostnames unless the operator intends that disclosure. Infrastructure metadata can be sensitive even when it is not a credential.

Consider DNS changes

A host validated at one time may resolve differently later. Send-time validation and address binding matter when a document is actually sent. Do that work off the main UI thread so network lookup does not freeze the app. Cancellation must close the socket rather than merely hide a loading indicator. Tests should distinguish policy rejection from cancellation and from remote errors.

A private address does not prove that the user owns the server. Shared networks, compromised hosts and misconfigured services remain possible. The product can enforce a destination class, but the operator still needs service identity, access controls and retention policies. For nodes, a pinned TLS fingerprint adds a separate identity expectation; for AI, the configured trust model must be described honestly.

Complete a transport exercise

Attempt a public AI endpoint and preserve the refusal. Start a legitimate loopback server, configure the reversed port and preserve discovery behavior. Stop the server and confirm the app reports unavailable rather than displaying an invented answer. Where HTTPS is used, test a wrong certificate or pin according to the module's rules. These results qualify bounded transport properties, not every possible network or every claim of private AI.

---

Wiki home · Qualification updates · Sources and credits