Sigilant Request a briefing
Why How it works Capabilities Deployment Request a briefing Sign in

Cintelis AI

Cintelis AI Other platforms

Sigilant · Control for AI agents

Every agent action, authorised.

AI agents are already raising orders, changing records and moving money inside your systems, on passwords nobody can trace to a person. Sigilant makes every agent act as itself, with permission, with proof.

Request a briefing How it works
Isometric diagram: requests from AI agents on the left pass through the Sigilant gateway in the centre, where each is allowed, refused or held for a person, and the allowed ones go on to business systems on the right.

Proudly partnered with

SCX.ai logo Sovereign inferenceSCX.ai DekkoSecure

Does this sound familiar?

Three questions you will be asked about your agents.

Each one has an honest answer today, and it is not a good one. Sigilant exists so the answer becomes a record you can show.

Three AI agents share one key into a business system

The purchase order (from our demo)

"An agent raised a $48,000 purchase order on Friday night. Who approved it?"

Today: the agent used a shared password. The finance system logged the password, not the agent. Ten agents use the same one, so nobody can say which agent did it, for whom, or whether anyone said yes.

With Sigilant: the order was over the limit, so it waited. A named manager approved that exact order, and the record shows her approval.

Three AI agents on a register, each with its own identity and a named owner; one is linked to the person it is acting for

The audit

"Show me your register of AI agents, and who owns each one."

Today: a spreadsheet, maybe, and agents running on a staff member's login. You are now expected to give each AI agent a unique identity, distinct from the user accounts of your staff and the identities of other AI agents, and to keep a register of its owner, its business purpose and what it can access, regularly verified.

With Sigilant: every agent gets its own identity: never a person's login, never shared. Each one has a named owner who confirms every so often that it is still needed. When an agent acts for someone, the record names both.

An AI agent stopped by a barrier in front of a locked business system

The rollout

"Security said no. The agent would need the finance system's password."

Today: the safe answer is no, so the project stalls. Or it goes ahead with a password hidden inside the agent, and everyone hopes.

With Sigilant: the agent never holds a password. Sigilant lets it in for the one action it is allowed, then closes the door. Security can say yes.

What it costs you today

Four things that are true until an agent has an identity.

01

You cannot answer the regulator

Who did it, on whose behalf, and who approved it. Three questions, no reliable answer.

02

You cannot stop it in time

Turning off a shared password breaks everything that uses it. So nobody does, until it is too late.

03

You do not know how many you have

Agents arrive inside software you buy and tools your teams build. The ones you cannot see are the ones that hurt.

04

You are the one saying no

Every agent the business wants is a risk you carry personally. Saying yes needs something to point to.

A real run

What governing one agent looks like, minute by minute.

From our demo: a purchasing agent, a manager called Priya, and a finance system. Nothing in the agent or the system was rebuilt.

  1. 1Priya logs in the way she always does. Her agent is now known as hers.
  2. 2Her agent raises a purchase order for $48,000.
  3. 3That is over the company's limit, so Sigilant holds it. The finance system has not seen it.
  4. 4Priya sees exactly what is waiting and approves that order, and only that order.
  5. 5Only now does the finance system accept the order.
  6. 6Later, someone switches the agent off. It stops at once, along with everything it started.
Console: the approval record for erp.create_purchase_order, approved by Priya for one exact call with a 48,000 limit
The approval, as the record holds it: which agent, for whom, approved by whom, for this call only.
Console: the audit chain with 204,856 records and 112 anchors, and the button to verify the whole chain
Every decision on one record that cannot be quietly changed. Your auditor can check it without taking our word for it.
Console: the agent register, with each agent's owner, sponsor, tools and whether it is still wanted
The list of agents: who owns each one, what it may touch, and when its owner last confirmed it is still needed.

How it works

One checkpoint. Four things happen at it.

Agents go through Sigilant to reach your systems, instead of going straight in. Neither the agent nor the system behind it has to be rebuilt.

Step 01

Decide

Each action is checked against your rules before it happens: which agent, for whom, doing what.

Allow, hold or refuse

Step 02

Approve

Sensitive or high-impact actions wait for a named person's approval before they run, and that person sees exactly what they are approving.

A person in the loop

Step 03

Access

The agent never holds the keys to your systems. Access is supplied for the one action, then gone.

Nothing to steal

Step 04

Record

Every decision, allowed or refused, is written to a record that shows if it has been altered.

Evidence as a by-product

What you can now say

Eight sentences for the board.

Each one is a capability, written the way you will have to say it.

  • Every agent has a unique identity and a named owner, who confirms every so often that it is still needed.
  • Nothing happens outside our rules. Each action is checked as it happens, not once when access was first given.
  • Sensitive or high-impact actions were approved by a named person before they ran, and the approval covered that action only.
  • No agent can do more than the person it works for, or more than its task needs: it gets the minimum set of tools and permissions required for its purpose.
  • No agent holds a password to our systems. It is let in for one action, then the door closes.
  • We can stop any agent, and everything it started, at once. Even halfway through a task.
  • Nothing an agent reads can override our rules, its permissions or a required human approval. A tricked agent still faces the same checks.
  • Our records can be checked by anyone, without trusting us or our supplier, and they are protected against tomorrow's code-breaking computers.
One agent cut off at the gateway while another continues
The kill switch
A chain of sealed records anchored to a vault
The record

Find them first

Find the agents you did not know you had.

AI agents arrive inside software you buy and tools your teams build, and most organisations only meet theirs during an incident. Sigilant looks first: it searches your code and your systems for AI agents, AI tools and passwords to AI services, and asks a person to claim each one or switch it off.

It keeps looking. Every change your teams make, every release and every day, it checks again, so "are we still safe?" has an answer you can show, not a feeling.

A lens over a cluster of agents, some registered and lit, some unknown and dark

Deployment

Built to fit what you already run.

Your people log in and approve with the login system you already use, such as Okta, Auth0 or Microsoft Entra ID, and agents belong to the teams you already have. Sigilant runs on your own systems: your rules, your records and what your agents do stay with you.

  • IdentityWorks with your existing logins
  • AgentsWorks with the AI agents you already use
  • HostingRuns in your environment
  • FailureIf it cannot decide, it refuses
  • CryptographyProtected against future code-breaking
  • EvidenceRecords nobody can quietly change

Bring one agent and one system it touches.

We will show you how Sigilant controls it, on your systems, with your rules. One hour, and you leave with the record.

Request a briefing How it works

Get in touch

Request a briefing.

Tell us which agents you are putting into production and what they need to reach. We will reply within one business day.