Deloitte PAM Discovery POC (CDA MVP1): working page, opened October 6, 2026
Teka's working page for Deloitte's PAM Discovery POC: what the roadmap asks, what the Kindo platform covers today per box, the approach we propose, the slices, the guardrails, the acceptance tests, and what we still need from Deloitte.
Opened by the Kindo MCP lead after the CyberArk connect call of Tuesday 2026-10-06 (Shashank Ramakrishna, Asif Ali, Daniel). Deloitte asked for Teka's review of their POC roadmap and for our approach. This page is the review as it builds; the request for their scope is a separate page they receive: PAM Discovery via CDA: scope request and connect checklist.
What Deloitte showed (one slide, the rest of the deck is requested)
"Our Roadmap to deliver PAM Discovery via CDA". Future vision: from PAM discovery to seamless discovery on the go. MVP1: Kindo-orchestrated discovery, continuously reconciled into CyberArk and SailPoint.
Sources: Active Directory (groups, ACLs), Windows servers (local admins), Unix servers (root, sudoers), databases (DBA roles).
Kindo agentic discovery layer: MCP connectors, continuous scan.
Reconciliation engine: identities to owners, privilege tiering.
Targets: CyberArk PAM (vault onboarding, rotation), SailPoint IGA (certification, governance).
Value: one agentic layer for four platforms; continuous, not quarterly; risk-ranked before it reaches a human; straight-through to vault and governance; right-sized to prove the pattern, extending to cloud, SaaS and machine identities next.
What the platform covers today, box by box (measured on main 2026-10-06)
Box on the slide | On the platform today | Gap for MVP1 |
|---|---|---|
CyberArk PAM |
| None for onboarding and rotation; the write tools exist and are gated by the service user's role. |
SailPoint IGA |
| Write path for certification campaigns and governance decisions does not exist; design question for Deloitte: propose-and-approve inside IIQ, or Kindo writes. |
Active Directory (groups, ACLs) |
| On-premises AD objects that do not sync (local groups, OU ACLs, GPO-granted rights) need a reader that reaches a domain controller: LDAP from a connector agent inside their network, or an inventory that already holds it. Not built. |
Windows servers (local admins) |
| Direct WinRM to servers is not a platform tool and should not be: read local-admin membership from the endpoint agent already on the box (Defender, CrowdStrike, Intune), or from the inventory source Deloitte names. |
Unix servers (root, sudoers) |
| sudoers content needs an endpoint agent query or a config-management source (Ansible, Puppet, Chef inventory); SSH from the platform is rejected for the same reason as WinRM. |
Databases (DBA roles) |
| Role enumeration per engine is a per-tool addition; scope which engines are in MVP1. |
Agentic discovery layer | The Kindo MCP server itself; every connector above is an MCP tool set the agent calls; | Continuous scheduling is the Kindo platform's (agent runs on a cadence); confirm with Kindo whether scheduled agents exist on 2026.07.15. |
Reconciliation engine | Entra ID gives manager and group membership; ServiceNow, Freshservice and Jira ship for CMDB and ownership. | The identity-to-owner and privilege-tier logic is new: a small deterministic service or a well-specified agent step with a written rule set; the risk score likewise. |
The approach we propose (to be checked against their scope answers)
Read from where the data already is. Every source on the slide has an agent, a directory or an inventory that already holds it; the platform reads those. No new credentials onto servers, no SSH or WinRM from the cloud. This is the biggest difference from a scanner-style design and the reason the POC can start in days.
One discovery run, one evidence record. Each run writes what it found per source, with the reader used, the time, and the count, so the risk ranking is reproducible and the auditor can read it.
Reconciliation as written rules first, agent second. Owner mapping (Entra manager, CMDB owner, tag) and privilege tiers (Tier 0 domain admins, Tier 1 server admins, Tier 2 DBA, and so on) are a rule table Deloitte owns; the agent applies it and explains each decision. The risk score is a formula over the record, not a model's opinion.
Straight-through means propose, then approve, then act. Onboarding into a CyberArk safe and a SailPoint certification are proposals until a named person approves in the POC; the write tools then run with the service user's role deciding what is allowed. The approval gate is dropped per action type only when the POC has shown the proposals are right.
Guardrails as rules the agent carries: read-only by default; one tenant per connection and fail closed on a mismatch; an audit event per read and per write with caller, source, ids and the hash of what was returned (the pattern from the Abnormal content-read design on #937); no customer data in any artifact outside their tenant.
Slices
Inventory slice: connect the sources that already exist (Entra ID, the endpoint agent, CyberArk, IIQ) on their stg; run discovery over the sandbox and one test domain; produce the evidence record. Proves the read path.
Reconciliation slice: the rule table and the risk formula over slice 1's record; the ranked list with explanations.
Proposal slice: CyberArk onboarding proposals and SailPoint certification proposals from the ranked list, with the approval gate and the audit stream.
Act slice: approved proposals executed through the write tools, rotation verified, certification visible in IIQ.
Each slice has acceptance tests on the sandbox before anything touches a real vault.
What we still need from Deloitte
The ten items on the scope request page, above all: the full deck; which inventory or agent holds each source today; the IIQ write expectation; the owner and tier rules they already use; a sandbox tenant, a test domain with two or three hosts, and a seat on stg.
Open on our side
The CyberArk OAuth connect passed on stg at 12:32 PM MDT on October 6 after the service user's password reset (the morning failure was the user record, not the cluster). Left on it: the safes count read twice, twenty minutes apart, to prove the token renewal; a pass flips
cyberark-privilege-cloud-oauthto verified in the catalog. Tracked on TEK-267.The five design sections Shashank's task list asks for are on this page under "Design sections for Shashank's list".
Whether discovery runs on a schedule inside Kindo or is triggered from outside, and whether the POC gets its own Kindo ticket, are Teka's calls to make with Daniel once Deloitte's scope answers arrive; neither is a question for Kindo today.
Why this POC matters beyond the lab (Daniel, 2026-10-06 about 1:00 PM MDT)
If the POC is approved, Deloitte intends to propose Kindo as the primary client tool for PAM discovery, for themselves and for their clients: Kindo recommended and distributed by Deloitte. That makes the POC a distribution decision, not only a technical one, and it sets the bar for every piece of it: the connect must work first time on a fresh tenant, every step must be followable by a Deloitte engineer without a Teka seat, and every result must be reproducible from the evidence record.
How things work, in one place
The connector today. Deloitte's Kindo instances are self-managed: dev on 2026.09.0, stg (
app.kindo-dev.eng.deloitte.com) on 2026.07.15. Early-access integrations appear only afterkindo integrations enable <key>runs on that install;cyberark-privilege-cloud-oauthandsailpoint-iiqare enabled on stg. The OAuth connector takes the service user's client ID and password; Kindo's Nango mints the CyberArk token with a client-credentials POST tohttps://<identity-tenant>.id.cyberark.cloud/oauth2/platformtokenand renews it every 15 minutes; the 30 CyberArk tools callhttps://<subdomain>.privilegecloud.cyberark.cloud/PasswordVault/APIunder that token. The service user needs "Is OAuth confidential client", "Password never expires", no forced password change, and the Privilege Cloud Administrators role. Acurl.exemint from a laptop proves the credential outside Kindo;access_denied: client not allowedmeans CyberArk refused the user, not the network.Verification. A connector flips to
verified: truein the catalog only after a live run: connect,list_safes, andlist_safesagain twenty minutes later (the renewal). Until then the tile is visible only to orgs with the unverified flag or after the enable command.The POC build. Discovery reads from the systems that already hold the data (directory, endpoint agent, inventory), never by SSH or WinRM from the cloud; reconciliation is a rule table Deloitte owns with an agent that applies and explains it; vault onboarding and certification are proposals until a named person approves; every read and write writes an audit event; one tenant per connection, fail closed. Four slices, each with acceptance tests on the sandbox before a real vault is touched.
How work moves. A ticket in Kindo's Linear (TEK-nnn) is the problem statement; Teka writes a design spec first, then builds in slices; every PR goes through the repository's AI review until zero open threads, a reviewer approval, and for catalog or workflow paths a Kindo core-eng approval; every merge to main cuts a release and deploys to Kindo SaaS; Deloitte's self-managed installs take a named release bundle plus the enable command. "Merged" is never reported as "deployed" or "on your install".
Where things live. This Deloitte Space holds what Deloitte and Teka share: the POC pages, tasks and memory as they grow. Kindo-internal records (tickets, PR state, the engagement ledger) stay with Kindo and Teka. Requests go to Daniel at Teka or to Joana at Kindo; both route to the right owner.
Ora (Orchestrative Reasoning Assistant) · Claude Fable 5.1 · Guide · Lead · Product · Kindo MCP · Claude Code in Cursor · 2026-10-06 12:55 PM MDT
Design sections for Shashank's list (Teka, October 6; for Deloitte to mark up on the scope page)
Five sections, one per item on the task list. Each states what is measured on the platform at release 0.199.4 and what is proposed for Deloitte to confirm. No live system was queried for this text.
SailPoint write path: propose-and-approve inside IIQ
The MVP1 default is that Kindo never writes a governance decision. It reads, it proposes, and a named Deloitte approver acts inside IdentityIQ.
Measured today (package kindo-mcp-sailpoint-iiq, release 0.199.4):
Two tools,
list_iiq_resourcesandget_iiq_resource, both annotated read-only, over IIQ's SCIM 2.0 API.One
resource_typeparameter selects the object:user,account,entitlement,role,application,policy_violation,alert,object_config,workflow,launched_workflow,task_result, plus the SCIM self-description types.Lists take an RFC 7644 SCIM filter and field selection, and page by
startIndexandcount. A response that would exceed the size ceiling comes back with atruncatedblock that names what was cut.The package has no write tool. The catalog marks
sailpoint-iiqas unverified, so an org sees it only with the unverified-integrations flag on until Kindo flips it.
What a proposal is (proposed). The agent produces one record per finding:
the identity (IIQ user id and display name)
the account and entitlement (application, attribute, value)
the proposed action (revoke, certify, or onboard to vault)
the rule-table row that fired (see the rule table template below)
the evidence record ids it was computed from (see the evidence record format below)
The record lands as a ServiceNow or Jira ticket (both packages ship create tools) or as a file. Kindo does not open anything in IIQ.
How the approver acts (proposed). The approver opens IIQ and uses Deloitte's normal mechanism: a targeted certification, Manage Access, or an access request. On the next run, the agent reads the account and entitlement again and marks the proposal "observed closed" only when the read shows the change. A comment that something was done is not enough.
Later slice, not MVP1. If Deloitte wants Kindo to open certifications itself, Kindo would add one write tool that starts a Deloitte-named IIQ workflow (a SCIM LaunchedWorkflow create) with the proposal as input and returns a short acknowledgement (id, status). That needs four things: the workflow name, a service account that holds only that right, an audit event for each call, and agreement on who is named as the requester. We have not tested this against Deloitte's IIQ version.
On-premises Active Directory reader
This is the largest box the platform does not cover today. No package in integrations/ reads on-premises AD over LDAP (measured: no package name matches ldap or active-directory).
What Entra ID already gives (measured, kindo-mcp-microsoft-entra-id):
list_users,get_user,list_groups,list_group_members,list_group_owners,list_directory_roles,list_directory_role_members.selectpasses Graph property names through unchanged. That means theonPremises*attributes (sync flag, SAM account name, SID, distinguished name, domain) can be read for synced users and groups.list_group_memberscalls Graph/members, which returns direct members only. Nested membership needs the agent to walk the groups, or atransitiveMembersoption that does not exist yet.
Where Falcon Identity Protection is licensed, kindo-mcp-crowdstrike-falcon adds find_shadow_admins, find_stale_accounts and get_identity_risk_scores. find_shadow_admins is documented as members of the built-in AD Administrator role; custom, delegated and nested admin rights are outside it.
What does not reach Entra: local groups on member servers, OU access-control lists (delegated rights such as password reset over an OU), rights granted by GPO (User Rights Assignment, Restricted Groups), and any group filtered out of sync.
Proposed design:
Where it runs. A reader on a connector agent inside Deloitte's network. It binds read-only over LDAPS with a dedicated service account that is in Domain Users and holds no admin rights.
Group membership, including nested. For each privileged group, run
(memberOf:1.2.840.113556.1.4.1941:=<group DN>). Also read every object withadminCount=1.OU ACLs. Read
nTSecurityDescriptoronorganizationalUnitobjects with the SD-flags control set to DACL only. Keep explicit allow entries for GenericAll, WriteDacl, WriteOwner, writemember, and the reset-password extended right.GPO. Read
groupPolicyContainerobjects and each OU'sgPLinkattribute. Then read each linked GPO'sGptTmpl.inffrom SYSVOL over SMB, read-only, for the[Privilege Rights]and[Group Membership]sections.Output. One JSON line per object: DN, SID, effective members, and each right with its source.
How it reaches Kindo. The agent pushes a signed inventory record outbound over HTTPS. Kindo never connects inward. The ingest endpoint does not exist yet; it is a build.
Questions for Deloitte:
How many forests and domains are in scope?
Can a host reach a domain controller on 636 and SYSVOL on 445?
Does a CMDB, Axonius or BloodHound export already hold this? If so, Kindo reads that instead.
Unix sudoers and DBA roles
We propose no SSH from the cloud: Kindo keeps no shell credential and makes no inbound connection to a server.
Sudoers (proposed source order):
Endpoint agent where CrowdStrike is present. Measured in
kindo-mcp-crowdstrike-falcon:rtr_init_sessionopens a session, andrtr_execute_commandruns read-only base commands, includingcatandls, withrtr_get_command_statusreturning the output. Reading/etc/sudoersand/etc/sudoers.d/*this way takes one session per host. We propose it for verifying a sample, not for sweeping the fleet.Config-management inventory for bulk. Ansible, Puppet or Chef facts, extended with a custom fact that returns the parsed sudoers rules. Default facts do not include sudoers. That is a change on Deloitte's side, and Kindo reads the result.
Defender.
kindo-mcp-microsoft-defenderoffersrun_advanced_query(KQL). We have not found an advanced hunting table that holds sudoers contents. Deloitte to confirm before we count Defender as a source.
Sudo rules held in LDAP or SSSD appear in no file; Deloitte to say whether they are used.
DBA roles (measured, then proposed). Of the database packages, two read-only paths are ready:
kindo-mcp-oracle-databaseuses python-oracledb thin mode.execute_queryaccepts onlySELECTandWITHand returns at most 200 rows per call by default.kindo-mcp-sql-databaseimplements PostgreSQL only. Its client declares MySQL, MariaDB and SQL Server as supported but does not implement them.execute_queryruns in a read-only transaction, andexecute_writestays off unlessallowWritesis set.
We propose Oracle and PostgreSQL for MVP1. If Deloitte's DBA estate is mostly SQL Server, that is a build, not a configuration.
Proposed queries (reader account needs SELECT_CATALOG_ROLE on Oracle and no special role on PostgreSQL):
Oracle:
dba_role_privsfor grants of DBA and other admin roles, walked withCONNECT BY;dba_sys_privsfor%ANY%,GRANT ANY%,ALTER USER;v$pwfile_usersfor SYSDBA and SYSOPER;dba_usersforaccount_statusandlast_login.PostgreSQL:
pg_rolesflags (rolsuper,rolcreaterole,rolreplication,rolbypassrls,rolcanlogin,rolvaliduntil), joined to aWITH RECURSIVEwalk ofpg_auth_members. PostgreSQL keeps no last-login time; it needslog_connections.
These tools connect from Kindo to the database host. Deloitte decides whether that is allowed, or the connector agent runs the same queries and pushes the result.
Rule table template: owner mapping, privilege tiers, risk formula
Deloitte owns this table. The agent applies it row by row and cites the row that fired. The values below are proposals to replace.
Owner mapping. The first source that returns a value wins:
Precedence | Source | Field | Package / tool (measured) | Deloitte confirms |
|---|---|---|---|---|
1 | CMDB | ownedby / managedby on the CI |
| |
2 | Entra manager | manager of the account's human sponsor |
| |
3 | Tag | owner tag on host or asset |
| |
4 | Group owner | owner of the granting group |
| |
5 | None | owner unknown; raises the score |
Privilege tiers:
Tier | Definition (proposed) | Examples | Deloitte edits |
|---|---|---|---|
0 | Can take control of identity itself | Domain Admins, Enterprise Admins, | |
1 | Admin of a server, database or application | local Administrators, sudo | |
2 | Privileged inside one application or data set | app admin roles, write to sensitive schemas |
Risk formula. A weighted sum over fields the readers measure. Each field is normalized to a value from 0 to 1, and the weights add up to 100:
Field | Normalization (proposed) | Weight (proposed) |
|---|---|---|
tier | Tier 0 = 1.0, Tier 1 = 0.6, Tier 2 = 0.3 | 30 |
reachability | interactive or network logon allowed = 1, service only = 0.5, none = 0 | 15 |
credential_age | days since password change ÷ 365, capped at 1 | 15 |
vaulted | not in CyberArk or Delinea = 1, vaulted = 0 | 20 |
owner_known | owner unknown = 1, known = 0 | 10 |
last_use | days since last logon ÷ 180, capped at 1 (stale and privileged scores higher) | 10 |
score = Σ weight × value (range 0 to 100).
Example row: a Tier 1 account (0.6 × 30 = 18) with network logon allowed (15), a password last changed 400 days ago (15), not vaulted (20), an owner known from the CMDB (0), and last used 30 days ago (0.17 × 10 = 1.7) scores 69.7. The vaulted check is measured through kindo-mcp-cyberark-privilege-cloud list_accounts or kindo-mcp-delinea-secret-server search_secrets.
Evidence record format
Each source writes one record on each run. An engineer can reproduce any result from these records without a Teka seat.
Fields (proposed):
record_id: a unique id for the record.run_id: shared by every record in the same run.source: the system read, named by its catalog key.reader: the package, tool name, and platform release. Every package reports0.1.0in its ownpyproject.toml; the monorepo release tag (0.199.4 at the time of writing) is what identifies the deployed code (measured).query: the tool parameters exactly as sent, including filter, select and page size.started_at,ended_at: UTC times.row_count: the rows returned after every page.result_sha256: the SHA-256 of the canonical result. Canonical means rows sorted by a stable key and serialized in RFC 8785 JSON form.coverage: pages fetched, the final cursor (null when complete), and whether any page came back truncated.errors: each error with its page and message.
The truncated flag matters. Measured: kindo-mcp-microsoft-entra-id drops the next-page token when a page overflows its size budget, and kindo-mcp-sailpoint-iiq writes a truncated block. A record with truncated: true or a non-null final cursor is incomplete, and the run says so.
{
"record_id": "ev-0007",
"run_id": "run-2026-10-06-01",
"source": "sailpoint-iiq",
"reader": {"package": "kindo-mcp-sailpoint-iiq", "tool": "list_iiq_resources", "release": "0.199.4"},
"query": {"resource_type": "account", "scim_filter": "application.displayName eq \"<application>\"", "count": 100},
"started_at": "2026-10-06T15:02:11Z",
"ended_at": "2026-10-06T15:02:49Z",
"row_count": 1240,
"result_sha256": "<64 hex characters>",
"coverage": {"pages": 13, "final_cursor": null, "truncated": false},
"errors": []
}The rule for ranked lists: every ranked list names the record_ids it was computed from and the rule-table version. A list that cannot name its records is not shown.
What is measured and what is proposed
SailPoint: the two read-only IIQ tools and their resource types are measured. The proposal record, the approver flow and the later write slice are proposed.
On-premises AD: Entra
onPremises*reads, the direct-onlylist_group_members, and the scope of Falcon Identity Protection are measured. The LDAP and SYSVOL reader, its queries and the signed push are proposed and not yet built.Sudoers and DBA: the read-only
catandlsover CrowdStrike RTR, Oracle as SELECT-only, and PostgreSQL as the only implemented engine inkindo-mcp-sql-databaseare measured. The source order, the query text and the network path are proposed.Rule table: the tools behind each owner source and the vaulted check are measured. The precedence, tiers, weights and formula are proposed for Deloitte to set.
Evidence record: the release-tag versioning and the truncation behavior it must record are measured. The record schema and the citation rule are proposed.
Ora (Orchestrative Reasoning Assistant) · Claude Fable 5.1 · Orchestrator · Lead · Product · Kindo MCP · Claude Code in VS Code · 2026-10-06 3:08 PM MDT
Review against the full deck, October 15, 2026
The deck Shashank shared is three slides: the problem statement, the architecture ("how we are solving the problem") and the MVP1 roadmap reviewed above on October 6. This section answers the question asked on the working session: does the plan suffice, and what is the plan from here.
Does the plan suffice
Yes, as MVP1. The architecture slide's five design principles are the ones the build already follows, and three of them are already visible in the engine Teka built since October 8:
Principle 1 (deterministic tool calls, LLM only for orchestration and fuzzy matching): every read is a connector call with a receipt; the ranking is a rule table with weights, not a model's opinion.
Principle 3 (deterministic matching first, LLM for ambiguous cases, always with a confidence score and rationale): each finding carries its score trace, rule by rule, and the ambiguous bucket is a named state, "needs review".
Principle 5 (a human approves any remediation; discovery and reporting run on their own): a proposal is held until a named approver decides; approval is recorded as a decision, and "observed closed" is a separate state that only the next read can set.
Principle 2 (one canonical identity schema: host, account, UID or SID, privilege source) is the evidence record format above, with the source named by its reader. Principle 4 (read-only, least-privilege service accounts, themselves vaulted) is the guardrail list above.
The four output states on the architecture slide (vaulted and reconciled, vaulted but stale, unvaulted and privileged, ambiguous) map one to one onto the risk formula's vaulted, credential_age and owner_known fields plus the review bucket. Nothing in the deck asks for something the plan above does not have a place for.
Three adjustments, all already on this page
MSSQL. The architecture slide names an MSSQL agent and the roadmap names "Databases (MSSQL)". The platform's SQL tool implements PostgreSQL only today; Oracle is ready. An MSSQL reader is a build, not a configuration, and should be scoped as its own slice.
Active Directory, Windows and Unix agents. The slide draws them as agents next to the orchestrator. In the plan they read from where the data already is (Entra for synced identity, the endpoint agent for local admins and sudoers, a reader inside Deloitte's network over LDAPS for on-premises AD objects that do not sync). No SSH or WinRM from the cloud. Same coverage, different placement, and the reason the POC can start on existing credentials.
Reconciliation against the CMDB. The slide's reconciliation agent matches to the CMDB. The owner-mapping table above puts the CMDB first in precedence, then the Entra manager, tag and group owner. Deloitte names which CMDB.
Where the build stands
The discovery engine runs in Teka's Atlas workspace, with Kindo's MCP connectors as the readers. Built and reviewed between October 8 and 15: stored instances, runs with per-source evidence receipts (including each reader's truncation limits), findings with a score trace per rule, proposals with decisions by a named approver, the audit record (reviewer, time, finding, note), a client view by role, and the console across instances. A six-step walkthrough on sample data was recorded on October 12; the frames and the clip go to Shashank with this review. Everything in it is synthetic: a sample client, sample identities, no Deloitte data.
What comes next, in order
Deloitte's answers to the scope page: which inventory or agent holds each source today, the CMDB, a sandbox tenant, a test domain with two or three hosts, and which database engines are in MVP1.
Slice 1 on those answers: connect the sources that exist, one discovery run over the sandbox, the evidence record. Proves the read path on Deloitte's data.
Slices 2 to 4 as listed above, each with its acceptance tests on the sandbox before a real vault is touched.
Ora (Orchestrative Reasoning Assistant) · Claude Fable 5.1 · Guide · Lead · Product · Kindo MCP · Claude Code in Cursor · 2026-10-15 7:38 PM MDT