Anoryx Sentinel
In development · recruiting design partnersThe self-hosted ownership control plane for every key, bot, AI agent and shadow AI app. Find them all, give each one an owner, and prove it to your auditor, without sending your identity data to anyone.
Identity Debt Score
live scanNon-human actors · 4,108
- Service accounts1,284
- API keys & tokens2,317
- OAuth apps412
- AI agents37
- Shadow AI apps58
AI agents and AI apps counted separately, never folded into “other”.
Top findings · by blast radius
0/5The problem · identity debt
Your machines outnumber your people. Most have no owner.
A person gets all five parts of an identity. A key, bot or agent usually gets the first three. When its creator leaves, it keeps working, and nobody answers for it.
The first question in every incident response: “What exactly do we have, and who owns it?”
01 · usually present
Actor
The thing that acts: a service account, a bot, an agent.
02 · usually present
Credential
How it proves itself: a key, token, secret or certificate.
03 · usually present
Permission
What it can reach: scopes, roles, policies.
04 · usually missing
Owner
The accountable human who answers for it.
05 · usually missing
Lifecycle
When it is reviewed, rotated and retired.
0:1
machine identities for every human
Source: CyberArk, 2025
0:1
once AI agents are counted
Source: Palo Alto Networks, 2026
0%
of ex-employees kept at least one working account
Source: Sagiss, 2026
How it works
Discover. Link. Own. Score. Fix. Prove. Repeat.
One loop that turns a pile of unknown credentials into owned, scored, evidenced actors, and keeps them that way.
Step 1 of 6
Discover
then repeat, every night
- 01
Discover
Read-only collectors sweep every source nightly and hourly.
Entra ID, Google Workspace, AWS, GitHub, OpenAI and Anthropic admin APIs, plus an endpoint script for MCP configs and AI extensions on laptops. Metadata only: secrets become fingerprints.
SHA-256 + last 4 · never the value
- 02
Link
Every object becomes one actor in one ownership graph.
A single schema (creator, owner, last used, scopes, expiry, source) across every ecosystem. Edges record who created it, who consented, what it uses and what it can reach.
created_by · owned_by · can_reach · acts_for
- 03
Own
Evidence becomes a suggested owner, then a confirmed one.
Creator lineage, sponsors, CODEOWNERS and tags combine into a ranked candidate with a confidence score. Owners confirm by e-mail in about two minutes (target), with no login. Unanswered requests escalate.
≥ 0.60 confidence → ask · else → orphan finding
- 04
Score
Named rules fire. One number tells you if debt is falling.
A YAML rule library with stable IDs mapped to the OWASP NHI Top 10, plus the Identity Debt Score: 0–100, a transparent weighted formula, no machine learning, recomputed daily.
unowned 0.30 · stale 0.20 · over-privileged 0.20 · …
- 05
Fix
The smallest safe fix, approved by a human.
Each finding ships with a guided runbook in the MVP. Approval-gated one-click revoke and rotate, with two-person approval, comes Next. Nothing is ever revoked automatically.
MVP: runbooks · Next: approved one-click
- 06
Prove
Evidence an auditor can read, from a log no one can rewrite.
The Identity Debt Report (PDF and CSV) and a hash-chained, append-only audit log with at least one year of retention. Then the loop runs again: nightly rescans, quarterly re-attestation.
hash-chained · append-only · ≥ 1 year
Architecture
A control plane, inside your boundary.
Ten steps from source to evidence. Sentinel calls admin APIs on a schedule and never sits in your traffic. Everything after the first read happens on your infrastructure.
Nightly
full sweep
Hourly
delta sync
0
secret values stored
0
bytes of telemetry out
Swipe to see the full flow →
Capabilities
Six jobs. One control plane.
Everything Sentinel does falls into six groups. Each line is tagged with its planned phase, so you know exactly what a design partner gets first.
Discover
Find every non-human actor, read-only.
- Entra ID / M365 incl. Agent ID sponsors, or Google WorkspaceMVP
- AWS multi-account incl. Bedrock keys; GitHub via a private AppMVP
- OpenAI and Anthropic admin keysMVP
- Endpoint scan for MCP servers and AI extensions on laptopsMVP
- Okta, Azure, GCP, GitLab, Salesforce, Atlassian, Copilot StudioNext
Own
Give every actor an accountable human.
- One actor schema and ownership graph with path queriesMVP
- Creator lineage from audit logs, stored from day oneMVP
- Suggested owners with evidence, never stated as factMVP
- Ownership queue and e-mail attestation, no loginMVP
- Slack attestation; optional local model for inferenceNext
Control
Score the risk. Decide, don't just ban.
- Rule library with stable IDs, mapped to OWASP NHI Top 10MVP
- AI agent registry: acts for, purpose, tools, scope, expiryMVP
- Shadow AI decisions: sanction, review, offer an alternativeMVP
- Findings with guided runbooksMVP
- Two-person approved one-click fixes, policies, simulationNext
Prove
Evidence an auditor can follow.
- Identity Debt Score: 0–100, transparent formula, no MLMVP
- Identity Debt Report as PDF and CSVMVP
- Hash-chained, append-only audit log, ≥ 1 year retentionMVP
- Signed, framework-mapped evidence packsNext
Leavers
When people leave, their machines don't stay behind.
- Leavers view from your HRMS CSVMVP
- Rule LVR-01: leaver credential still activeMVP
- Reassign owner, rotate copied secrets, revoke grants, record evidenceNext
Stays home
Runs in your network. Answers to you.
- Docker Compose on one VM in your VPC or data centreMVP
- No phone-home; you pull updates; mail via your relayMVP
- Connector credentials in OpenBao or your cloud KMSMVP
- Helm, HA Postgres, signed images and SBOMNext
Console tour
See the console before it ships.
Concept screens with a fictional tenant. Click through what an admin, an owner and an auditor would see.
Every non-human actor in one table, across every source. Saved filters find the debt in one click.
Risk rules
Named rules, not a black box.
Every finding cites a rule with a stable ID, mapped to the OWASP Non-Human Identities Top 10. Nine ship in the MVP; eighteen are planned in total.
# rules/nhi-01.yaml
id: NHI-01
when: owner is null or owner.status == left
severity: by blast_radius
fix: runbook/assign-owner
Orphaned
No accountable owner, or the owner has left.
Stale
Unused for longer than your threshold, still valid.
Over-privileged
Holds far more permissions than it uses.
App-only high-risk scope
Acts with no user, on sensitive scopes.
Unregistered agent
An AI agent with no sponsor, purpose or expiry.
MCP with prod credentials
A laptop MCP config holds a production key.
Unsanctioned AI app
An AI app nobody approved, holding company data.
Broad OAuth scopes
Drive-wide or mailbox-wide access via consent.
Leaver credential active
Someone left. Their key still works.
Next: NHI-05–08 · AGT-02 · AGT-04 · SAAS-02 · SAAS-04 · LVR-02
The trust contract
Every byte stays home.
Sentinel is a control plane, not a gateway. It keeps the inventory and the rules. It does not store secrets, sit in the login path, broker sessions or read what people write.
Reads
- Admin-API metadata via read-only scopes, listed line by line
- Application.Read.All · Directory.Read.All · AuditLog.Read.All
- AWS SecurityAudit · a private GitHub App
- Key fingerprints: SHA-256 plus the last four characters
Never reads
- Secret values. Discovery credentials cannot write.
- E-mail bodies, files or chats
- Prompts or model outputs
- Live traffic. It is not in the request path.
Nothing leaves
- Runs inside your VPC or data centre
- No outbound telemetry, no phone-home
- You pull updates; attestation mail uses your relay
- Connector secrets in OpenBao or your cloud KMS
Deployment modes
| Mode | Shape | Phase |
|---|---|---|
Pilot One VM, Docker Compose, about 8 vCPU / 32 GB. Install within a day (target). | One VM, Docker Compose, about 8 vCPU / 32 GB. Install within a day (target). | MVP |
Production Kubernetes via Helm, HA Postgres, RBAC, signed images and SBOM. | Kubernetes via Helm, HA Postgres, RBAC, signed images and SBOM. | Next |
Regulated / air-gapped Offline bundle, local model, no egress at all. | Offline bundle, local model, no egress at all. | Roadmap |
MSSP One instance per client, with a partner console. | One instance per client, with a partner console. | Roadmap |
Every line above is a design commitment for a product in development, not a description of shipped software. Sizing figures are targets, not measurements.
Who it's for
One question each. One place to answer it.
Built for regulated and security-conscious mid-market companies, 200 to 5,000 people, rolling out AI agents and copilots. India first.
CISO / CTO
“How much identity debt do we carry, and is it falling?”
One score, its trend, the top five findings.
IT / security admin
“Who owns this key, and can I retire it?”
Lineage, blast radius and a suggested owner.
Owner / engineer
“Are these still mine?”
One e-mail. Keep, transfer, not mine, retire.
Head of AI
“Is my agent registered and scoped?”
A registry with sponsor, tools and expiry.
DPO / compliance
“Which apps can touch personal data, and who approved them?”
Shadow AI decisions with scopes and owners.
Auditor
“Who or what can reach this system, and who approved it?”
Path queries and a tamper-evident log.
HR
“Has the leaver's machine footprint been handled?”
Every key, token and agent tied to a leaver.
Typical triggers: an ISO 27001, SOC 2 or RBI / SEBI CSCRF audit, DPDP readiness, a leaked key, an agent rollout, or a senior engineer handing in notice.
Honest edges
What Sentinel is not.
A control plane earns trust by being clear about its limits. Sentinel works alongside your IdP, vault and security stack. It does not replace them.
Not a vault
It never stores secret values. It keeps fingerprints so it can recognise a key, and leaves storage to your vault.
Not inline, not a proxy
It does not sit in the traffic, the login path or a session. It does not block requests, inspect content or redact prompts.
Not a CASB or AI firewall
It finds shadow AI apps through OAuth grants and endpoints, then helps you decide. It does not police traffic.
Not auto-revoke
Nothing is revoked or rotated without a human. MVP fixes are guided runbooks; one-click fixes (Next) need approval.
Evidence
Maps to the frameworks your auditor asks about.
Sentinel produces audit-ready evidence for access-control and identity clauses: who owns what, who approved it, and when it was last reviewed. Framework packs are planned for Next.
Audit-ready evidence, never certification. Sentinel produces evidence; it does not make anyone compliant. Anoryx holds no certifications today.
Help shape Sentinel.
Become a design partner.
We're recruiting a small group of regulated, mid-market teams to shape the MVP: a first scan on your own infrastructure, a direct line to the people building it, and a roadmap that answers to your audit.