Solution Paper
VPNs and TLS Protect the Pipe, Not What AI Providers Keep: Glossary
Plain meanings of the terms used in VPNs and TLS Protect the Pipe, Not What AI Providers Keep.
Allowlist. A list of the only destinations something is permitted to reach. The opposite of a blocklist, which names what is forbidden and permits the rest. An allowlist is the stronger of the two and, as the paper shows, still not sufficient on its own: if an allowed destination is somewhere an attacker can also write, the allowlist has not closed the path.
Attestation. The mechanism by which a machine proves, cryptographically, what software it is running. The important question is never whether attestation happens but who holds the relying party role: the party that inspects the evidence and decides whether to proceed. When that party is the provider, or a service the provider runs, the proof is weaker than it sounds.
Attestation report (versus a certification). In standards language these are different products. A certification means an accredited body audited an organization against a published standard and issued a pass. An attestation report is an auditor’s opinion about controls the organization itself selected and described. Both are useful. Only one of them means somebody else set the bar.
Bind address. The network address a server listens on. 127.0.0.1, called loopback or localhost, means only the same machine can reach it. 0.0.0.0 means every network interface, which on a machine with a public address means the internet. The difference is one line of configuration and it is the single most consequential setting in a self-hosted deployment.
Business associate agreement. Under United States health privacy law, the contract a covered entity must have in place with any vendor that handles protected health information on its behalf. A model provider touching patient data needs one. The obligation attaches to the vendor relationship, not to the technology.
Confidential computing. Hardware features that encrypt a running workload’s memory and exclude the machine’s own operator from reading it, so that even the host operating system and the provider’s administrators see ciphertext. Genuinely strong against software adversaries. The vendors publish a list of what it does not cover, and the paper quotes that list.
Controlled Unclassified Information (CUI). A United States federal classification for information that is sensitive but not classified, with a published category list and handling requirements. Public bodies and their contractors have these categories handed to them and should use them rather than inventing a parallel scheme.
CVSS. The Common Vulnerability Scoring System, the 0 to 10 severity number attached to published vulnerabilities. Useful for triage and, as the paper shows, systematically misleading for this category of flaw, because a score is reduced when an attack requires privileged local access even where that privileged local party is precisely the adversary the defeated product existed to stop.
DNS. The system that turns a name like api.example.com into a numeric address. Two things make it matter here: it is a channel data can be smuggled out through, and its queries are frequently absent from the network logs an organization believes are complete.
Egress. Traffic leaving a network. Egress control means deciding what may leave, to where. Exfiltration is the hostile case: data leaving that should not have.
Enclave. A protected region of a computer in which code and data are shielded from the rest of the machine, including from privileged software. The building block of confidential computing.
Indirect prompt injection. The central unsolved security problem in this field. An AI system reads some content, a document, a web page, an email, a code comment, and instructions hidden in that content are followed as though the user had given them. It is called indirect because the attacker never talks to the system; the attacker only has to get something in front of it. No reliable prevention mechanism exists, which is a statement of industry and government consensus, not an opinion.
Inference server. The software that loads a model and answers requests against it. The thing you actually install when you self-host. Several popular ones are named in the paper by their behavior rather than their brand, because their defaults differ in ways that decide whether a deployment is exposed.
Least privilege. The principle that any component gets the narrowest access that lets it do its job, and no more. Applied to AI agents it means read-only by default, scoped credentials per agent, and never a shared administrative credential.
Mesh overlay. A private network built peer-to-peer across the public internet, so that machines in different places reach each other without either being exposed to the internet. The right tool for reaching your own self-hosted systems. Each product has a coordination or trust dependency worth naming before adopting it.
Private endpoint. A network arrangement that gives you a private address inside your own network for a cloud service, so traffic to it never traverses the public internet. Removes public exposure. Does not change the fact that the service receives your data in readable form.
Prompt. Everything you send the model: your question, plus any document, file or context attached to it. The reason a “short question” can carry a whole contract.
Redaction. Removing or masking sensitive values before data leaves. Deterministic redaction means rules you wrote, in code you control, over categories you defined, as opposed to asking a model to decide. The paper’s position, which rests on the detection vendors’ own published disclaimers, is that redaction is worth doing and must never be the only control.
Reverse proxy. A server that sits in front of another server and handles the connection on its behalf. The place to terminate authentication, because it can cover every path, whereas an inference engine’s own key may cover only some.
Reasonable measures. The statutory phrase in United States trade secret law requiring the owner to take reasonable measures to keep the information secret. It is part of the definition of a trade secret, not a defense, so failing it means there was never a protected asset rather than merely a weaker case.
Side channel. An attack that learns secrets not by breaking the encryption but by observing something incidental: timing, power draw, cache behavior, memory access patterns. Explicitly out of scope for every confidential computing product on the market, and the subject of the 2026 research the paper cites.
Single tenancy. Hardware or an instance not shared with other customers. Removes co-tenants as an adversary. Does nothing about the provider, which is a different question and the one most buyers are actually asking.
Split tunnelling. A VPN configuration in which only some traffic goes through the tunnel and the rest goes straight out. Common, often sensible for performance, and it quietly removes the traffic-inspection benefit organizations believe they are getting from the VPN.
Subprocessor. A third party your vendor hands your data to in order to deliver the service. The list of them, and the notice you get before it changes, are contract terms worth reading rather than assuming.
TLS. Transport Layer Security, the encryption behind the padlock in a browser. Version 1.3 is current. It protects data from anyone watching the path between you and the service, and provides exactly no protection against the service itself, which is the destination rather than a party on the path.
Trusted computing base (TCB). The set of components you have no choice but to trust, because if any of them is compromised your protection is gone. The vendors’ own definitions place firmware and microcode inside it, which is why vulnerabilities in those have the impact they do.
Work product. In United States litigation, material prepared in anticipation of litigation, which is protected from disclosure separately from attorney-client privilege. The paper cites a 2026 decision holding that a defendant’s conversations with a consumer AI assistant were neither.
Zero data retention (ZDR). A contractual arrangement under which the provider does not store your inputs and outputs. Real, available from several providers, and in every case a gated product rather than a setting: it requires prior approval and comes with a list of excluded features. Worth asking for by name and worth reading the exclusions of.
Back to VPNs and TLS Protect the Pipe, Not What AI Providers Keep
Find out what your project really takes.
Book a free 30-minute consultation. You will leave with a realistic sense of size, cost and risk, whether or not we work together.