Authority and scope
Who authorized which targets and operations? Which version of the rules applied at run time?
A record that work was delegated to an AI internally and permission obtained from the target organization are different things.
VAP application example · Verifiable Evidence for AI-Assisted Security Research
Trace who authorized the work, what the agent observed, and how each finding was reached.
This application proposal combines VAP observation records and delegation records so that whoever receives a finding can see what it rests on and how far it has been confirmed.
“Outdated plugin detected.”
What it rests on, and how far it was confirmed, cannot be seen.
Which part of which response was matched against which information — with the conditions that remain unconfirmed.
Synthetic data No cryptographic verification has been performed. See the full example
When an AI reports that it has found a problem, that report alone leaves four things unclear.
Who authorized which targets and operations? Which version of the rules applied at run time?
A record that work was delegated to an AI internally and permission obtained from the target organization are different things.
How was each URL discovered? What was actually retrieved?
From which URL, when, and by what route the material behind a judgment came is part of the record.
What was observed, and what was inferred? If an assessment was corrected, can the original judgment and the reason be traced?
Unless observed facts, referenced information and unconfirmed conditions are kept apart, inference travels as if it were fact.
Do the allow/deny decisions match the traffic that actually took place?
Whether decision and execution matched can only be checked by comparing both records.
The record that research was authorized, the pre-retrieval decision, the actual traffic, the technical assessment, and the decision on notifying the target company are each a different record.
For example, a record that an operation was decided as “allowed” does not show that the operation was carried out as decided. Conversely, a record of traffic alone does not show whether that traffic stayed within what was authorized.
VAP (Verifiable AI Provenance Framework) is a cross-domain framework for making records of AI decisions and actions verifiable after the fact. VAP v1.2 registers OAP, which covers observation of web material published by third parties, and DAP, which covers actions an agent takes under delegated authority. Both are Working Drafts.
Covers the provenance of web material that a third party published and that was retrieved over a network, observed at a URL. It targets records of material that may later be disputed, deleted, edited or denied.
Covers the provenance of agent access, tool invocation, action and data transfer under authority delegated by a principal. It is not tied to any domain and is used together with a domain profile.
Both inherit the requirements of the Shared Assurance Core below
Requirements inherited by every profile and capability, including the integrity, provenance, traceability and accountability layers, external anchoring (INT-006), and the Completeness Invariant (INT-008), which makes post-anchor omission and split-view presentation detectable at anchor granularity.
For security research, we are considering linking the following information to observation records and delegation records. Which record should carry each item, and in what form, has not been decided.
The left side of each row is what VAP v1.2 states; the right side is this page’s own application proposal. The right side is not a definition in OAP, DAP or VAP.
OAP is a domain profile for the provenance of records of web material published by third parties and observed at a URL (§2.2.6, §5.4). It refers to an ObservationScopeManifest and to an attempt-closure invariant, Invariant OAP-1 (§5.4; their definitions are in the OAP text).
Link how each URL was discovered, and the basis for each product and version observation, to the observation record.
DAP is a cross-cutting capability for the provenance of agent access, tool invocation, action and data transfer under authority delegated by a principal (§4.6.2). VAP’s responsibility schema includes a delegation_chain of delegator, delegatee, scope and validity period (§4.4.3).
Treat internal delegation to an agent and permission obtained from the target organization as separate records.
A deployment consists of one domain profile and zero or more declared capabilities; where constraints overlap, the stricter applies (§4.6.3). OAP composes with DAP where retrieval is performed by an autonomous agent (§4.6.4, §5.4).
Link OAP and DAP records along six research stages: setting the scope, delegation, pre-retrieval decision, observation and retrieval, technical assessment, and notification, correction and handover.
VAP’s policy identification (policy_id, §7.1) identifies the verification policy under which events were produced. It does not represent the rules for what research may or may not do.
Reference the version of the research rules (targets, permitted and prohibited operations, period of application) from the delegation record and the pre-retrieval decision record.
External anchoring is required at all conformance levels (INT-006). For any anchored batch, a third party must be able to detect omission or split-view presentation after anchoring (Completeness Invariant, INT-008). Events that were never recorded are out of reach (Appendix A).
When verification is implemented, state the trusted record source, how records are fixed externally, and the scope being verified.
There is an abstract provenance model (actor, input, context, action, outcome; §4.2.2) and an override_history (§4.4.3). The VAP v1.2 text defines no records specific to security assessments or notification decisions (the OAP and DAP texts have not been checked).
Treat observed facts, referenced information, likely applicability, unconfirmed conditions, the notification decision and the reason for a correction as separate records.
Conformance to VAP or a profile does not constitute compliance with any law and does not warrant that recorded decisions are correct, fair or safe; it concerns only the integrity, completeness at anchor granularity, and attributability of the records (§1.6). VAP does not prevent, block or intercept AI behaviour (Declaration).
Design the evidence mechanism and execution control as separate mechanisms.
The stages below are an application proposal for linking records (proposed on this page). Under “related specification elements”, each stage lists only what VAP v1.2 actually states.
Within the organization doing the research: who entrusted which agent with what scope of work. Changes to and revocation of the delegation are recorded too. The intent is to map this to what DAP covers — actions under delegated authority (proposed on this page).
Which scope and period of testing the target company has authorized. This is a separate record, linked to the version of the authorization letter or contract.
Internal delegation creates no authority to access third-party systems. Operations that presuppose the target organization’s permission (for example, additional testing under a permission or contract) are linked to the record of that permission; an internal delegation record is never used in its place.
Execution control enforces the scope of operations in the system.
The evidence mechanism produces records of those decisions and of their execution that can be checked afterwards.
Evidence records alone cannot prevent prohibited operations. VAP is a framework for after-the-fact verification and does not prevent, block or intercept AI behaviour (VAP v1.2, Declaration). Preventing prohibited operations requires separate execution control.
An example built from a fictitious target and synthetic data. It contains no real company, URL or vulnerability information. The display is static and for explanation only; no signature or hash value has been verified.
| Data type | Synthetic data |
|---|---|
| Target | A fictitious public website, https://shop.example/ (.example is a domain name reserved for examples) |
| Origin of the URL | A reference in public HTML (a script referenced by the home page’s HTML) |
| Rules applied | Demo research policy demo-policy, version 3 (fictitious) |
| What was observed | A plugin name and an apparent version string, demo-gallery 2.3.1 |
| Matching result | Matches the affected range “below 2.4.0” of demo advisory DEMO-ADV-0001 (fictitious) |
| Unconfirmed | The version of the code actually running, whether individual patches were applied, whether the feature is in use |
| Assessment | Conditional problem candidate |
| Notification status | Not sent |
| Cryptographic verification | Not performed in this example |
All record IDs are fictitious values for explanation. They are not official OAP or DAP event names or field names. Opening this list does not verify any signature, check any external anchor, or detect tampering.
If detection of alteration, omission or substitution is implemented in future, at least the following three points will be stated (proposed on this page).
Whose records, signed with which keys, serve as the starting point for verification.
In which outside service (a timestamp authority, a transparency log, and so on) record digests are fixed, and how often. In VAP v1.2, external anchoring is required at all conformance levels (INT-006).
Which batches, periods and kinds of record were verified. Post-anchor omission or split-view presentation can be within reach of detection; actions that were never recorded cannot be detected (VAP v1.2 §4.1.7, Appendix A).
VAP v1.2 recommends an Evidence Pack: a bundle a third party can verify offline, containing events, audit paths, anchor records, external anchor proofs, public keys, policy documents and similar material (§9.3). The example on this page is not an Evidence Pack.
These are the premises for evaluating this evidence design. Read them as part of the design.
| What the design aims to do | What that alone does not guarantee |
|---|---|
| Detect alteration of stored records | That observations represent reality completely and accurately |
| Trace judgments and their basis | That a vulnerability exists, or that it can be exploited |
| Cross-check delegation records against operation records | That the target company gave permission, or that the research was lawful |
| Verify evidence within a declared scope | That no unmeasured or unrecorded action took place |
Adding a hash afterwards does not, by itself, show that an approval record existed before the research. This design considers linking approval records and execution records to external anchors, so that one can check by when each record existed. The approver’s authority and the validity of the approval need to be confirmed separately.
This proposal is an evidence design for verifying research records after the fact. It does not grant access to any target, and it does not guarantee that a vulnerability exists, legal compliance, or independent certification.
Position in the specification: conformance to VAP or a profile does not constitute legal compliance and does not warrant that recorded decisions are correct (VAP v1.2 §1.6). Events that were never measured are, by construction, outside the reach of verification (VAP v1.2, Appendix A).
EVIDECT is a security research service being developed by VeritasChain Co., Ltd. We are considering a design that links observation, delegation, assessment and notification records to the process of examining public information and similar sources for possible problems.
Support for VAP, OAP and DAP as presented on this page is an application design proposal. Implementation status, test results and any conformity assessment will each be shown together with material that can be checked.
No. Verifying the evidence can show whether records were altered afterwards and how a finding links back to the original observations and judgments. On its own, this design does not guarantee that a vulnerability exists or that it can be exploited.
Both are Working Drafts registered in VAP v1.2.0 Draft 3. OAP (Observed Artifact Provenance) is a domain profile for the provenance of records of web material published by third parties and observed at a URL. DAP (Delegated Access Provenance, common name VAP-AGENT) is a cross-cutting capability for the provenance of agent access, tool invocation, action and data transfer under delegated authority.
Internal delegation creates no authority to access third-party systems. This design treats internal delegation and the target organization’s permission as separate records. It does not grant access to any target, and it does not guarantee that research is lawful.
No. VAP is a framework for after-the-fact verification and does not prevent, block or intercept AI behaviour. Preventing prohibited operations requires separate execution control that enforces the scope of operations in the system.
No. EVIDECT is a service being developed by VeritasChain Co., Ltd., and its evidence integration using VAP, OAP and DAP is at the design stage. No conformance test results or independent verification results have been published.
Write to the standards contact listed in the VAP specification, standards@veritaschain.org. Comments do not amount to normative changes to VAP or any profile.
The material this page refers to, and the publication status of related outputs. Status is always given in words.
We welcome comments on this application design, for example on:
Comments do not amount to normative changes to VAP or any profile. Specification changes follow VSO’s change management process (VAP v1.2 §10.3).