VerifAIdentity

Platform

Seven capabilities. Three working today.

VerifAIdentity is an assurance layer for identity: it checks that identity controls took effect in every system, and keeps the evidence. Here is the whole plan, and exactly how far each part has got.

The seven pillars

Built in lab works on lab data in our own tenants In development partly built Planned not started
  • 1

    Cross-vendor sign-in monitoring

    Built in lab

    Sign-in and admin events from every provider, normalised to OCSF and tied to one person, so an attack that is quiet in each system is visible in one place.

    Today: Okta System Log, Entra audit log and PingOne activities; seven detections with published thresholds. Entra sign-in logs need Entra ID P1 or P2.

  • 2

    Identity assurance: compliance clocks and gap checks

    Built in lab

    Leavers, privileged access, access versus approval, inactivity and credential expiry run as deadline clocks with per-system evidence, layered access and explicit gaps.

    Today: five controls, six clock states, access layers AL-1.0, signed evidence chain EC-1.0, audit pack and OCSF export.

  • 3

    AI identity analyst

    Built in lab

    Plain-English questions answered only from the connected data, citing records, declining when the evidence is not there.

    Today: rule-grounded Ask and Daily Brief; 40-case evaluation (34 answered, 6 declined, all correct). The model-backed analyst is the next step.

  • 4

    Non-human identity governance

    In development

    Workloads typed and owned: service principals, managed identities and service accounts, with their secrets and their owners.

    Today: identity types IG-1.0, credential expiry clocks, Microsoft first-party apps separated. Owners are collected from the next Entra sync; lifecycle is not managed.

  • 5

    AI-agent access verification

    In development

    AI agents as first-class identities: who owns them, what they can reach, and whether delegation from a person can be shown.

    Today: agents are typed by Entra tag (medium confidence) or name (low confidence). The lab has no AI agents yet, and delegation is not observable in the connected systems.

  • 6

    Identity verification

    Planned

    Checking that the person behind an account is who HR says, at hire and at risky moments, with proofing results kept as evidence.

    Today: not started. Would connect to identity-proofing services rather than build one.

  • 7

    Behavioural assurance

    Planned

    Session-continuity signals that show whether the person using a session is still the person who started it.

    Today: not started. Needs privacy review, accuracy targets and customer consent before any build.

Nothing here is production ready. "Built in lab" means it runs end to end on our simulated company with automated tests; a pilot is how each part gets measured on real data.

One graph for people, workloads and AI agents

Not every account belongs to one person. Each account gets a type, an owner state and the reason for both, with a confidence level.

Identity graph A person with three accounts and a governance record; a workload owned by the person with a secret; an AI agent with an owner and an unobservable delegation. approved forownsowns (not yet collected) acts for? David Okaforperson P006 Okta accounthuman, high confidence Entra accounthuman, high confidence Governance identityrecord, not a sign-in account HR worker recordE1006, the trigger source Approved: Okta, Entra accessfrom governance AI agent (example)typed by tag or name, with confidence Workload (service principal)owner: recorded, none or vendor secret, expires Feb 12
Solid lines are observed in the connected systems. Dashed lines are relationships the model supports but the lab cannot observe yet. The lab has no AI agents; the agent node shows how one is typed.
  • Types: human (member, guest, break-glass), workload (service principal, managed identity, service account), AI agent, application, HR record
  • Owner states: a person, one or more recorded owners, none recorded, the vendor, or not collected yet
  • Unmatched accounts stay unmatched as orphans; nothing is forced onto a person
  • Reconciliation always shows its denominators: in our lab, 12 of 27 people are in two or more sign-in systems and 15 in one; 69 of 69 human accounts matched

The controls checked today

Each rule is versioned and has an owner, a severity and a scope. Deadlines are your settings, not ours.

ControlWhat has to be trueDefault deadlineNot checked yet
Leaver access removed everywhereEvery sign-in account in scope is off after the HR termination date; each account described in access layers.24 hoursApplication-level sessions; policy-defined exceptions
Privileged roles approvedEvery admin role holder has a matching approval in governance.72 hoursApproval authority, duration and PIM activation history
Access matches approvalEvery active sign-in account traces to an approved access record.72 hoursEffective permissions inside applications
Inactive accounts disabledEach person is used or disabled within the period, judged across all their accounts.90 days, warning 14 days beforeSystems without sign-in data (shown as cannot be verified)
Credentials rotated before expiryApp secrets and certificates are rotated before their end date.Warning 30 days beforeSecrets outside Entra app registrations

Clock states

Every clock is in exactly one of six states.

  • Breached

    The deadline has passed and fresh evidence shows the control still failing.

  • Cannot be verified

    A source is stale or does not report what the rule needs. Shown as a gap, never as a pass.

  • Running

    The clock has started and the deadline has not passed yet. You see how long is left.

  • Fixed late

    Fixed after the deadline. Kept as an audit exception with how the fix time was found.

  • Met

    Fixed in time, confirmed by fresh evidence from every system in scope.

  • Retired

    The rule or account no longer applies. Kept in history and in the evidence chain, not deleted.

Cross-vendor detections

Thresholds are settings, shown here at their defaults. Each detection lists the telemetry it needs; test cases and known limits are on the Methods page.

DetectionFires whenNeedsMITRE ATT&CK
Password spray across systemsOne source address fails against 5 or more people within 30 minutes, in any connected systemFailed sign-ins with source addressT1110.003
Brute force against one person8 or more failures for one person within 15 minutesFailed sign-ins per userT1110.001
Success after repeated failures3 or more failures, then a success, for one person within 60 minutesFailed and successful sign-insT1110, T1078
One person failing in several systemsFailures for the same person in two or more systems within 30 minutesSign-ins from two or more providersT1110
Activity by a person who has leftA sign-in or admin action after the HR termination dateHR export and sign-in or audit eventsT1078, T1078.004
Admin change by an unknown identityAn admin action by an identity with no governance recordAudit events and governanceT1098
Sign-in failure spikeFailures in one system reach 5 an hour and three times the normal rateA history of sign-in eventsT1110 (or an outage)

See it on your own identity data.

A read-only pilot with Okta, Microsoft Entra ID and your HR export. 30 to 60 days, measured against targets we agree up front.