| Internet-Draft | A Disconnected Trust Model for Networks | November 2026 |
| Morrison | Expires 22 May 2027 | [Page] |
The architecture document for IP in deep space says that security designs must tolerate long validation delays and potentially operate in a disconnected trust model, and it does not say what that model is. This document defines it. A disconnected trust model is one in which a node makes each trust decision from data it already holds or is handed with the request, each item verifiable without asking a third party, with a declared limit on how old each kind of held data may be. The document then asks what a node may do alone when it cannot reach an authority, and lists the requirements any answer must meet. It leaves short credential lifetime, as the answer to revocation without contact, with the documents that already give it. On the direction of failure, it adds that an authority check that cannot be evaluated fails closed on authority and never on safety. It specifies no protocol and no mechanism.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 22 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Section 10 of [TIPTOP-ARCH] says that security designs "must tolerate long validation delays and potentially operate in a disconnected trust model". Section 7 of [TIPTOP-UC] lists what follows from the delays. Online status checks such as OCSP take too long, certificate renewals must be planned ahead of the delay, and recovery from compromise needs remote means. Section 6.2 of the same document adds that a node's clock uncertainty grows while it has no time reference. Neither document defines the disconnected trust model.¶
Four documents already state part of its premise. Section 10 of [TIPTOP-ARCH] says certificates need to be validated using local trust anchors. Section 10 of [DIEM-ARCH] describes an offline bundle that a disconnected validator checks with a root trust anchor and a clock. Section 5.11 of [BPSEC-COSE] describes how a long-delay enclave can avoid online status checks. Section 2.1 of [BP-SAFE] describes keys that are created ahead of their use, each with a validity interval. A reader who asks what a node may do alone, holding a credential and unable to reach the party that issued it, finds a partial answer in each and no shared vocabulary.¶
This document gives the model a definition (Section 3), points at the answers that already exist (Section 4), names the questions those answers leave open at the authority layer (Section 5), and lists the requirements any answer must meet (Section 6). It is an Informational complement to Section 10 of [TIPTOP-ARCH] and Section 7 of [TIPTOP-UC].¶
It specifies no protocol, no message format and no mechanism. It makes no change to IP, and it does not choose among answers. Routing, transport, relay storage and group key management are out of scope.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
In this document the key words apply to an answer to the questions in Section 5. A "MUST" says that an answer which omits the point fails the requirement. No implementation conformance is defined.¶
The scenario labels CS, LF, LI, MF, MI and SC are those of Section 4 of [TIPTOP-UC].¶
A host, router, relay or enforcement point that makes a trust decision.¶
A party whose statement a node relies on to make a trust decision. A certificate issuer, the holder of a trust anchor, and the party that grants a node or an agent the right to act are all authorities.¶
Data a node stores and uses without asking another party. Trust anchors, credentials, status information and configuration are held data.¶
Data handed to a node with the request the node is judging.¶
The greatest age, as a duration, at which a given kind of held data may still be relied on. The deployment declares it.¶
A node is out of contact with an authority when it cannot reach that authority at the time of a decision. The term describes a pair of node and authority. A node can be out of contact with one authority and in contact with another.¶
The evaluation of whether the party making a request holds the right to make it.¶
An authority check is unevaluable when an input it needs is missing, malformed, older than its staleness bound, or depends on a clock the node cannot trust.¶
An action a system takes autonomously to protect itself or people, as described in Section 6 of [OT-COMMAND-AUTHORITY].¶
A disconnected trust model is a trust model in which every decision a node makes is made from held data and carried evidence, each item of which the node can verify without asking a third party, and in which the deployment declares a staleness bound for each kind of held data.¶
The model constrains where a decision's inputs come from and how old they may be. It says nothing about how the data reach the node, and it is neutral about the cryptography used to verify them.¶
Verification without a third party means that the node checks a signature chain to a trust anchor it already holds and checks the time against a clock it has reason to trust. [DIEM-ARCH] describes the same property for DNSSEC data in Section 10. An RRset, its signature and the chain up to a trust anchor can be validated anywhere by anyone holding the root trust anchor and a clock.¶
Carried evidence can be buffered at intermediate nodes for a long time before it arrives [TIPTOP-ARCH]. A node judges it at arrival, against its own held data.¶
Every authority check depends on held data of some age. A trust anchor, a credential and any information about the status of either has a staleness bound declared by the deployment. [OT-COMMAND-AUTHORITY] requires a deployment to declare its revocation latency budget and its maximum trust-anchor staleness, and says that it does not fully resolve the trade between revocation freshness and offline operability (Section 8). This document sets no values. In setting the bound, a deployment states how it weighs freshness against the ability to operate alone.¶
An authority check is satisfied, refused or unevaluable. A check that cannot be evaluated has not been refused by an authority and has not been satisfied. Section 4 of [OT-COMMAND-AUTHORITY] treats the unevaluable case as a refusal for authority, and Section 6 of this document states the same point as a requirement.¶
The model needs a time reference, and a node can lack one or hold one of growing uncertainty. [TIPTOP-UC] Section 6.2 describes holdover as the time until a node's clock uncertainty exceeds what its applications can tolerate. [RFC9171] Section 4.2.7 recommends that a node without an accurate clock set a bundle's creation time to zero. Section 4.4.2 of that document then requires the bundle to carry a Bundle Age Block, an age summed by the nodes that forward the bundle. That age is an input those nodes write. [DIEM-ARCH] Section 10 notes that a validator whose clock can be set back will accept expired signatures. In a disconnected trust model, therefore, the time a check relies on is trusted, uncertain by a stated amount, or absent.¶
Short credential lifetime is the existing answer to revocation without contact, and it belongs to the documents that give it. [BPSEC-COSE] Section 5.11 notes that OCSP enables prompt revocation but entails online access, and offers short-term auto-renewed certificates as the alternative across long-delay networks. Intermediate certification authorities place the issuing authority near the end entities it serves, and a validator then needs no online access to a responder. [DIEM-ARCH] Section 11 says that profiles needing prompt removal should use short signature validity periods and TTLs, and that this limits the shelf life of offline bundles. Section 7 of [TIPTOP-UC] says the renewal lead time should exceed the worst-case one-way delay, including expected disruptions, plus the growth of clock uncertainty over the renewal interval. Section 10 of [TIPTOP-ARCH] asks for renewal before expiry. This document offers no competing answer.¶
[BP-SAFE] Section 2.1 gives each content key a single validity interval and allows a sequence of keys with future validity to be created in advance. Rollover then needs no exchange at the time. Section 1.1 of the same document leaves certificate provisioning and revocation lists out of scope.¶
[DIEM-ARCH] Section 10 says an offline bundle is valid for no longer than the shortest signature validity period it contains. The same period bounds how long a removed emblem can be replayed. Section 11 separates the validity of an issuer's statement from the lifetime of a copy and says that only the latter governs removal.¶
These answers say how long a credential or key stays usable without contact and how a copy ages. They say less about the decision a node makes in the interval before a short lifetime ends, when the authority has changed its mind and the node has not heard. They also leave open what happens when the absence outlasts every lifetime the node holds. Those are the questions in Section 5.¶
Each requirement applies to any answer to the questions in Section 5.¶
Every input to an authority check MUST be held by the node or carried with the request, and MUST be verifiable without asking a third party.¶
A deployment MUST declare a staleness bound, as a duration, for each kind of held data an authority check relies on. The declaration MUST be available to those who review the node's decisions. Where the same data feed two checks under different bounds, each reliance is its own kind of held data, and the declaration names the bound and the check it governs.¶
Held data older than its staleness bound MUST be treated as unavailable. A check that needs it is unevaluable.¶
A node MUST NOT accept a request on the ground that its authority check could not be evaluated. An unevaluable authority check yields a refusal. This is the rule of Section 4 of [OT-COMMAND-AUTHORITY]. An authority check evaluates the inputs the request's class requires. An input the check does not require cannot make it unevaluable, and a profile that lets a class proceed without an input says so. Section 5 of [OT-COMMAND-AUTHORITY] makes the grant optional for its lowest class, Observe.¶
A safety function MUST NOT depend on the outcome of an authority check, on a staleness bound, or on the availability or freshness of any input held for an authority check. A safety action that a system would take autonomously MUST remain possible when every authority is unreachable. This is Section 6 of [OT-COMMAND-AUTHORITY].¶
A profile built on the model MUST state, for each action it covers, whether the action is gated on authority (R4) or is a safety function (R5). Where refusing an action is itself the harmful outcome, the profile MUST say so. Section 10 of [DIEM-ARCH] asks a validator for a trustworthy clock and does not say what the validator does without one. A profile answers that question for its own actions.¶
A node MUST state its time source and the uncertainty it assigns to its clock. A check that depends on a clock the node cannot trust, or on a timestamp that is absent or zero, MUST be unevaluable. A node MUST NOT treat a missing or zero time as the current time.¶
A profile MUST distinguish how long an issuer's statement is valid from how long a held or carried copy of it may be relied on (Section 11 of [DIEM-ARCH]).¶
The means by which a node's held data are replaced or withdrawn MUST work over arbitrary delay and intermittence. It MUST NOT need both ends to be online together, and a late or repeated update MUST NOT corrupt the node's held state. Section 7 of [TIPTOP-UC] asks for the first two of key update.¶
When a node has accepted a request, dispatched the action and cannot learn whether it took effect, the node MUST record the outcome as unresolved. It MUST NOT report the outcome as performed or as refused. An unresolved outcome MUST NOT carry authority forward to a retry, and a retry is a new request. This is Section 4 of [OT-COMMAND-AUTHORITY].¶
A node MUST keep a record of each authority check, naming the held data used, their ages and the staleness bounds in force. The decision can then be reviewed when contact resumes.¶
A profile MUST state its staleness bounds for each connectivity scenario of Section 4 of [TIPTOP-UC] that it claims to serve.¶
The requirements apply in every scenario, and the scenario changes the bound a deployment declares and how often a check is unevaluable.¶
Contact is mostly continuous, and intermittence is the exception ([TIPTOP-UC] Sections 4.2 and 4.4). Bounds can be short. R4 to R7 decide what happens in the exceptions.¶
The one-way delay grows slowly. The renewal lead time of Section 7 of [TIPTOP-UC] grows with it. The bounds declared under R2 are revisited as the delay changes.¶
Contact comes and goes. Section 4.5 of [TIPTOP-UC] gives an average of one contact every six hours for a surface asset served by a few orbiters. R9 matters most here.¶
Long-haul links are interrupted for a period of weeks, while links local to the body continue ([TIPTOP-UC] Section 4.6). A node out of contact with a distant authority can still be in contact with others. Questions 2 and 3 of Section 5 arise here more often than elsewhere.¶
This example uses invented values. A node N holds a trust anchor for authority A (authority.example). The deployment declares a staleness bound of 21 days for the anchor and 7 days for information about the status of credentials A has issued. N's clock has a stated uncertainty of 2 seconds.¶
N receives a request that carries a credential signed by A. Ten days after N's last update from A, the anchor is within its bound. The status information is 10 days old and is beyond its 7-day bound. It is treated as unavailable (R3). The check that needs it is unevaluable, and N refuses the request (R4). N records the check, the 10-day age of the status information and the bounds in force (R11).¶
The same node is asked to take a protective action that it would take on its own criteria with every authority unreachable. N takes it (R5). The profile in force says so for that action (R6).¶
N then receives a request whose creation time is zero, from a sender without an accurate clock ([RFC9171] Section 4.2.7). The maximum-age check on the request needs its creation time, and that time is zero. The check is unevaluable (R7), and N refuses the request. The bundle's age block does not repair it under R1, because the nodes that forwarded the bundle wrote that age and N cannot verify it. The credential's own validity is checked against N's clock and is unaffected.¶
This document specifies no mechanism. The considerations below apply to any answer that meets the requirements.¶
A node whose clock an attacker can set back accepts expired signatures ([DIEM-ARCH] Section 10). R7 makes the clock an explicit input with a stated uncertainty and makes a check that depends on an untrusted clock unevaluable.¶
A signed copy stays acceptable until its own validity ends, whatever the issuer has since decided ([DIEM-ARCH] Sections 10 and 11). A longer staleness bound widens the window in which a revoked statement can be replayed. R2 requires the bound to be declared so that the window can be judged.¶
R3 and R4 turn loss of contact into loss of authority once a bound passes. An attacker who can keep a node out of contact for that long can deny it the ability to accept authority-gated requests. [TIPTOP-ARCH] Section 10 describes flooding a link for a whole contact window. R5 limits the effect to authority-gated actions, and a deployment has to decide whether that loss is tolerable. A profile that answers question 3 of Section 5 with authority that wakes when contact is lost makes withheld contact a trigger as well as a denial. Whoever can withhold the data that would end the absence can then wake that authority. Such a profile names who can withhold those data, and states what bounds the authority they can wake.¶
A node that fails open on authority when a check cannot be evaluated accepts any request for which the attacker can make the check unevaluable. Under R4 the node refuses. A node that gates a safety function on an authority check turns an unreachable authority into a safety hazard. R5 forbids it.¶
Reporting an unresolved action as refused asserts that nothing reached the process, and the assertion may be false. R10 forbids it.¶
Carried evidence buffered at intermediate nodes is exposed if the node is compromised ([TIPTOP-ARCH] Section 10). A profile should say how long carried evidence may sit in a buffer.¶
Physical access to a deep space asset is generally impossible ([TIPTOP-UC] Section 7). R9 requires that the means of replacing held data work asynchronously.¶
Carried evidence that names a person, an agent or the party an agent acts for can sit in intermediate buffers for a long time before it arrives. Anyone with access to a buffer can read it. A profile should carry the least identifying evidence the check needs.¶
The record required by R11 holds who asked for what and which held data the node relied on. It is retained for review after contact resumes, and a deployment should bound its retention and its readers. [DIEM-ARCH] Section 11 observes that the DNS is not forgetful and that passive collectors may retain records indefinitely. Data placed in such a record should be assumed to persist.¶
Held data about authorities and credentials can reveal which parties a node trusts. A profile should state who can read a node's held data.¶
This document has no IANA actions.¶
draft-morrison-tiptop-disconnected-trust-00 (2026-11-18): first version.¶
None yet.¶