AgentAlign · security for AI agents

See your AI agents the way an attacker sees them.

We map which AI agents and machine identities can reach the most critical assets in your organisation, close the paths with the greatest impact, and stop actions agents should never take. Progress is documented monthly to management and the board.

Book an Agent Exposure Snapshot See the demo

Abstract digital network representing AI agent attack paths

Established partnerships

Why this approach

What makes it different from what you already have.

Most organisations already invest in Microsoft Defender, Entra ID, and SIEM. AgentAlign is not a replacement — it is the layer those tools cannot cover.

7 min

was all it took an AI-orchestrated attacker to delete over 100 storage accounts in Azure, via a single machine identity with overly broad permissions.

Microsoft Threat Intelligence, Sept. 2026

38 %

of Nordic organisations plan to deploy AI agents, while the share prioritising security has dropped from 59 to 42 percent.

Microsoft Work Trend Index Nordics 2026

1 year

since the Norwegian Digital Security Act came into force. Boards are responsible for knowing and managing digital risk — including the risk created by agents.

Digitalsikkerhetsloven, in force 1 Oct 2025

No unified viewAgents across Copilot Studio, cloud, code and SaaS — with no single picture of what they can reach.
No named ownerService accounts and agents without a named responsible person, so no one can revoke access.
Permissions too broadOne leaked key or one deceived agent opens a path to the most critical assets.
Control stops at loginA legitimate agent can cause damage within the access it already holds.

Same email · two outcomes

An AI assistant can empty your customer registry in seven minutes. Nobody has to break in.

A fictional company, Nordvik Trade, uses an AI sales assistant. One morning an email arrives with a hidden instruction. Here is what happens with and without AgentAlign.

Without AgentAlign

09:14:00An email arrives with a hidden instruction.
09:14:02The assistant reads it and does what it says.
09:14:05It exports the customer registry, using its admin account.
09:21:0048,210 customer records sent to an outside address.
No alarm. Every action used approved access.
Day 9A customer reports fraud. The leak is discovered.

0customer records exposed

72 h breach notificationGDPR fines up to 4 % of turnoverBoard accountabilityLost customer trust

With AgentAlign

09:14:00The same email arrives.
09:14:02The assistant tries to export the registry.
09:14:02The gate checks: Kari is not allowed to export, the recipient is external. Blocked.
09:14:03The owner in IT Security is alerted, with a full log.
09:14:03Business as usual. The assistant keeps working.

0customer records exposed

Why tools miss it

Nobody broke in. The assistant logged in correctly and used access it had been given. Firewalls, antivirus and login controls see nothing wrong. The risk is not who gets in, but what the agent is allowed to do once inside.

Why it happens · how we close it

Think of the AI assistant as a new employee with a master key.

Most organisations give their AI agents broad access to get started quickly. AgentAlign closes the gap in four measured steps, without stopping the agents from doing their job.

Step 1SeeWho has keys to what?
Step 2CutReplace the master key
Step 3GuardCheck every action
Step 4StopLock it in seconds

The assistant has a master key, and nobody owns it.

The challenge

To move fast in the pilot, the assistant was given an admin account. “We’ll tighten it later.” Later never came.

Why it’s a risk

Anyone who can talk to the assistant, or just email it, can open every door. Nobody notices, because the key is valid.

What we do

We map every AI agent, which doors its key opens and who is responsible. In 1–2 days, not months.

4 of 4 doors open · no owner

AgentAlign · how it works

Simulated · synthetic data

Step 1See
Step 2Cut
Step 3Guard
Step 4Stop

Paths to critical zone–

Exposure score–

Time to revocation–

Open pathCut or controlledBlocked by gate

See this on your own agents in 1–2 days.

An Agent Exposure Snapshot shows 3–5 concrete paths from your agents to one critical asset. Fixed price.

Book a Snapshot Technical demo →

Built for the Nordics

Your attack map stays home.

A map showing where an attacker can go is the most sensitive document an organisation has. That is why the service is built around three principles.

Data in the Nordics

Analysis runs in your own environment or hosted in Norway or the EU. We do not retain the map after delivery.

One register, multiple requirements

Each agent gets an owner, purpose, permissions and risk class, linked to the Digital Security Act, NSM's core principles, GDPR and DORA.

Employees protected

Logging of what agents do on behalf of employees is pseudonymised, with a fixed retention period and a template for consultation with employee representatives.

Offers

Start small. See the problem yourself.

The first step takes one to two days and reveals concrete paths from your own agents to a critical zone. From there, you decide whether it is worth going further.

Half a day · fixed price

Agent Exercise

For leadership team or board
Scenario: an AI agent is taken over
Who decides, how quickly is it stopped, and what is reported?

Start here

1–2 days · fixed price

Agent Exposure Snapshot

Initial mapping against Entra and Active Directory
3–5 concrete agent paths to one critical zone
Short report and walkthrough with CISO

3–4 weeks · fixed price

AgentAlign Assessment

Full graph of agents and machine identities
Critical zones and top-10 actions with owner
Exposure score and agent governance plan

Founding partners pilot

Ongoing · monthly

AgentAlign Continuous

Continuous mapping and real-time control
Owner and deadline per attack path
Monthly exposure report to management and board

Why StairPoint

We turn your Microsoft investment into measurable risk reduction — and cover what lies outside.

Cross-stack

We show paths that run through Microsoft, cloud, code and SaaS — not just within a single vendor's environment.

Independent

We see your environment the way an attacker does, without selling you a platform.

Accountable

We own the delivery: a named owner per agent, a deadline per attack path, and measured progress every month.

Founding partners – winter 2026/27

Three founding-partner slots. Open until 31 January 2027.

For organisations that have AI agents in production or pilot and want to know what those agents can actually reach — before someone else finds out.

Snapshot of one critical zone, with no commitment to more
Assessment and three-month pilot at a fixed founding-partner price
Direct access to the founding team and priority on the roadmap

Contact

Jørn Ellefsen

CEO, StairPoint Security · LinkedIn ↗

jorn.ellefsen@stairpoint.com

Send email about Snapshot