When AI Agents Can Act, Security Must Govern Their Actions
Useful autonomy depends on clear boundaries, approval for sensitive actions and evidence that people can review.
Give every team an AI workforce.
Ask a question across your selected servers. Find the cause of an outage, investigate an alert, collect the evidence for an audit. See the commands that ran, with changes governed by your host policy and human review when required.
Free for 7 days on up to 3 hosts, read-only by default. Run a sample investigation
Question
Which hosts still allow SSH password sign-in?
Agent investigates
Three hosts selected. Read-only session; every command is logged.
Systems and tools
$ sshd -T | grep -i passwordauthon srv-web-01 and srv-app-02; the macOS laptop gets its own check.
Evidence
srv-web-01: off. srv-app-02: on, set in sshd_config.d/50-cloud-init.conf. ops-laptop-07: still collecting.
Answer and action
One host still accepts passwords. The fix is proposed, not applied: it waits for your approval.
- Runs on your infrastructure
- Read-only by default
- Every command and source shown
- Arabic & English
How it works
- Connect
Your servers, documents and systems, and the model you choose.
- Ask
In Arabic or English, the way you would ask a colleague.
- Check and approve
Every answer shows its evidence. Review proposed changes; execution follows the configured host policy.
Knowledge
Policy
Systems
Files
APIs
Tools
Try it on your own servers, free
7 days, up to 3 hosts, one workspace for your team. A metered hosted model is included, or connect your own.
- Create your workspace
Sign up and your workspace opens on AcTree’s hosted service in Riyadh.
- Enrol up to three hosts
One install command and an enrolment token per host. Read-only by default.
- Ask
In Arabic or English. Read every command, its exit code and its output, per host.
Start free trial
Prefer a guided start on your own infrastructure? Book a technical briefing
The trial runs on AcTree’s hosted service: use non-sensitive servers and documents.
Watch it work
See all use casesThree of thirteen short films — the question, the steps, the evidence and the approval. All figures are sample data.
Built for IT, security and GRC teams
See all use casesThree teams get the most from AcTree today. Each question below is one a team would ask; every answer comes with the commands and output behind it.
IT operations
Find the cause, not just the symptom: every server checked the same way, every command shown.
- Which hosts miss security updates, have TLS certificates expiring within 30 days, or disks over 90%?
- Why is this service failing on one host since the last deploy?
- Which servers are missing updates or running out of disk?
Security
Investigate alerts and harden hosts with the evidence attached, read-only by default.
- Investigate this alert across the log platform, the endpoint tool and the host.
- Which hosts still have an active account for a departed employee?
- Which of these servers still permit SSH password authentication?
GRC and compliance
Collect control evidence from every server for your NCA ECC, PDPL and SAMA CSF work.
- Collect evidence for our access controls, and review the policy for gaps.
- Show the evidence for these controls on every production server.
Evidence
Evaluated, not just demonstrated
What we can show today, with its date. No customer quotes, logos or certifications appear here until they exist.
Test suites at the last release evaluation
In the 2 October 2026 release evaluation, every platform test suite passed: console 1,194 tests, gateway 970, portal 563, inference 48, host agent 42. The counts change as the suites grow; these are that day’s.
One scenario, end to end, on real hosts
The investigation on this site, SSH password sign-in across a three-host lab fleet (two Linux servers and a macOS laptop), was run read-only by the agent against those hosts. Its commands, exit codes and findings are what the demo shows.
Failures are reported, not hidden
On 2 October 2026 a host-agent release candidate was held back: it enrolled and reported in, but its runtime failed on a hardened Linux host. It ships when that passes, and the result stays on record.
Every claim traced to code
20 documented product capabilities on this site each reference supporting code. Site checks validate those references.
Private by design.
With Sovereign deployment, AcTree, your model, your documents and your servers stay inside your boundary. Nothing crosses it unless you choose an external model or service.
- On-premises
- Private cloud in Saudi Arabia
- Air-gapped
Customer-controlled inference and execution boundaries.
Customer-hosted deployments use packaged deployment configuration. An offline deployment requires an accepted bundle, local model and disconnected installation tests.
Where it runs is your decision
Three ways to deploy AcTree. The answers below change with the option you choose, so we show all three wherever we talk about deployment.
| Question | SovereignPlatform on infrastructure you control | HybridHosted control plane; your model, inside your network | HostedAcTree’s hosted service in Riyadh |
|---|---|---|---|
| Where AcTree runs | Entirely on infrastructure you control: your data centre or your Saudi cloud tenancy. | The control plane is hosted by AcTree; the host agents and any gateway run inside your network. | AcTree’s hosted service in Riyadh (stc cloud); host agents run on the machines you enrol. |
| Where inference runs | A model you run locally, or a provider you choose. | Your model, inside your network, through your own AI gateway. | The metered hosted model, or a model you bring. Prompts go to that model’s provider. |
| Where documents and evidence live | Your database and your storage. | Conversations and evidence in the hosted control plane’s database; documents stay with you unless you attach them. | AcTree’s database in Riyadh. |
| What can leave your boundary | Nothing, unless you choose an external model or service. | Prompts, command output and evidence, to the hosted control plane. Model traffic stays inside. | Prompts, command output and attached documents, to the control plane and to the model provider. |
| Who is in control | Your administrators, your identity provider, your keys and backups. | Your policies, approvals and model; AcTree operates the control plane. | Your workspace admins set policy and approve changes; AcTree operates the platform. |
| Air-gapped option | Yes: an offline bundle with a local model. | No: the control plane needs a connection. | No. |
In every option the agent is read-only by default. Commands follow the configured host policy: allowed, denied or referred for human review.
Architecture in detailSecurity & compliance
How AcTree supports your NCA ECC, PDPL and SAMA CSF work.
- NCA ECC
- PDPL
- SAMA CSF
Guides
Guides for buyers and engineers
- Private AIManaged private AI agents for IT and securityWhat managed private AI agents include: deployment, operating responsibilities, support scope and evidence for IT and security teams.
- Saudi AIPrivate AI in Saudi Arabia, explainedWhat "private AI" means for a Saudi organisation, why regulated teams ask for it, and how to run AI agents inside a boundary you control.
- Private AIHow to compare private AI agent servicesCompare hosted and customer-hosted agents by data flow, controls, operating responsibility, evidence and total cost.
- Private AIIT operations: review a host investigation with evidenceA sample IT workflow showing how to review host commands, exit codes and output, then measure a real pilot.
Insights
Explore insightsAI agents need clear boundaries.
Every agent should have a defined purpose, limited access and an accountable owner.
Sensitive actions should require approval, and completed work should leave evidence that people can review.
These principles guide our approach to AI agent security at AcTree.
Six weeks, on your infrastructure, with success criteria agreed up front.
- Week 1Install
AcTree runs inside your environment with one command, or as an air-gapped bundle.
- Week 2Connect
Your servers, documents and the model you choose. Read-only by default.
- Weeks 3–5Use it
Your teams work real cases in Arabic and English, with the evidence shown every time.
- Week 6Decide
We review results against the criteria you set, and you decide what happens next.
Pricing is quoted per deployment, after the technical briefing.
Questions buyers ask
Where does our data go?
It depends on the deployment you choose, and we show all three side by side. Sovereign: nothing leaves your environment unless you choose an external model or service. Hybrid: prompts and evidence go to the hosted control plane; your model stays inside your network. Hosted: prompts, command output and attached documents go to the control plane in Riyadh and to the model provider.
Which models can we use?
Bring your own model — OpenAI-compatible endpoints, Ollama, Anthropic, or Cloudflare AI Gateway / Workers AI — with keys envelope-encrypted at rest.
Does it work in Arabic?
Yes. You can ask questions and receive answers in Arabic. Website content is available in Arabic and English; some technical guides are available only in English.
What does the pilot need from us?
A server or VM to install on, read access to the systems and documents in scope, and two or three people who will use it.
What happens after the pilot?
If it met the criteria, an annual licence starts on terms agreed in the proposal. If it did not, you have the results report.
Can the agent change our systems?
Read-only execution policy is the default on every enrolled host; the full policy is an explicit, per-host choice.
Execution follows the configured host policy. A command can be approved, denied or referred for human review; individual approval depends on that policy.