Freezers set to +50 °C in several hundred stores, through the normal remote access. Every command was valid for the device.
OASIS by MandGard
Critical infrastructure that stays safe, even after a breach.
OASIS is the last line of defence against hackers and malicious AI agents. Every command is sealed and checked before the equipment executes it, the way a bank secures a transaction.
Book a live demo Send a test command
- Software only
- Equipment and programs unchanged
- Outside the safety chain
- Works without internet
Try it
Send a command. See where it stops.
OASIS, opened up: three barriers and a sealed record between any sender and your equipment. Pick a command and follow it.
Pick a command and watch where it stops. Select a step to see what it does.
Example: a cold store kept at −20 °C, limits set by its operator.
Proof
Eleven attackers at once, and no hostile command got through.
- 11
- automated attackers at the same time, six of them firing bursts from a valid operator account.
- 200,000+
- hostile attempts: out-of-limit values, forged or revoked senders, replays. None executed.
- 23,537
- legitimate commands during the attack, all executed, each one matching the record.
- 204,230
- entries in the record, refusals included; verified intact afterwards.
Our own measurements on one unmodified commercial logic module: single run, scripted attackers, no certification. Reproduction by a third party is planned for Q4 2026 to Q1 2027. All tests completed.
The problem
Attackers no longer break in. They log in.
Once an attacker holds a valid account, every command they send is obeyed. Firewalls, encrypted links, passwords and monitoring guard the way in or raise the alarm afterwards. Nothing checks the command itself.
First documented use of commercial AI models against an industrial target. The AI singled out a water utility’s control platform on its own. It stopped at a login prompt.
Frontier AI models met all eight attack objectives in an industrial test lab. From the perimeter to the control network: just over 16 minutes.
Introducing OASIS
Others guard the way in. OASIS guards the equipment itself.
OASIS is software placed in front of the controllers you already run. It lets a command through only when it carries a valid seal, stays inside your limits and, for critical actions, has a second approval. It applies to commands what a bank applies to payments: limits, identity, dual approval, no replay and a ledger.
A new seal for every command
Same operator, same command, same equipment: a different seal each time, valid only once. A captured command cannot be sent again.
Limits you set
Values stay inside your limits and change only in the steps you allow. Outside them, a command is blocked, whoever sends it and however many people approve it.
A second approval
Critical commands need two different people. One person alone, or the same person twice, is not enough.
A sealed record
Every attempt, passed or blocked, is kept, and any later edit shows. A stolen identity is revoked: its seals stop working.
We do not claim “unbreakable”. OASIS is built so that breaking one thing is never enough.
How it works: three steps, your controllers and supervision stay as they are
OASIS becomes the only path
It is placed on the link between the supervision and the controller, during a short planned intervention. Removing it restores the original link.
You set the rules
Limits and step sizes for each setting, which operators may act, and which commands need two approvals.
Every command is checked
Before it reaches the equipment. What follows the rules passes; the rest is blocked. Everything is recorded and can be verified by a third party.
Twelve properties, and what each one brings
Speed, memory and power cuts: our own bench measurements; speed measured against a simulated device. Today: one standard industrial protocol and one commercial controller measured.
Sectors
Versatile by design: one core, and everything else plugs in.
Water, energy, industry, pharma, cold chain: wherever software sends commands to equipment. Each new block opens a new sector; the checks and the record come with it, unchanged.
Any sender
Any equipment, any protocol
Industrial controllers are in the product today, over one standard industrial protocol; the other blocks are on our roadmap. Send us your controller and supervision models and we confirm compatibility before any pilot, with no commitment.
Compliance
Measured against the rules that arrive in 2027.
From 20 January 2027, the EU Machinery Regulation requires new equipment to resist tampering and to keep evidence of interventions. It covers new equipment only. OASIS brings the same kind of check to equipment already installed, and gives builders of new equipment a record of interventions on the command path.
EU Machinery Regulation, Annex III 1.1.9
Published test criteria for the EU machinery rules met on our bench; none failed, 3 not applicable. Tampering detected 3 of 3.
IEC 62443-4-2
Selected component requirements, from identification to the protection of audit information, met by system integration in our rehearsal.
Published attack patterns, 2023–2026
Replayed on our bench and blocked: exposed web panel (dam, Norway, 2025), hijacked operator screens, native protocols, stolen engineering keys.
Our own verdicts on our own bench. OASIS is not certified, and no presumption of conformity exists yet for Annex III 1.1.9. Evidence of hardware tampering is not covered. Each attack case is our own replay, not a reproduction of the original incident.
All tests completed on our bench, on a real commercial controller
| Test | Result |
|---|---|
| 11 automated attackers at once | Result200,000+ hostile attempts, none executed; no legitimate command lost |
| Two-person rule under attack | Result481 critical commands executed; 1,008 bypass attempts refused |
| Foreign, malformed and flood traffic | ResultNothing answered, nothing reached the controller |
| Slow drift inside the allowed range | ResultCapped: 17 changes accepted, 49 refused; 66 pass without the cap |
| 100 forced crashes, 4 real power cuts | ResultRecord intact after every restart |
| Two-person approval from two computers | ResultExecuted with two approvers; refused with one, or when replayed |
| Copied key | ResultA copy of the sealed key does not open elsewhere |
| Record edited on purpose and at random | Result5 deliberate edits and 320 random changes: all detected |
| Same commands sent without OASIS | ResultThe controller alone accepts them: the blocking comes from OASIS |
Our own measurements on one unmodified commercial logic module; several repeated three times; no certification, not yet reproduced by a third party. A forced crash is not a power cut; the drift test shows a speed limit, not prevention. The evidence file is shared under a non-disclosure agreement.
Pilot
A 30-day pilot, on a test bench or a chosen line.
OASIS sits between the supervision and the controller, without modifying the program of either. The pilot repeats our bench tests on your equipment.
Scoping
With your integrator: equipment in scope, limits, every access to the controller, what happens while OASIS is unavailable.
Installation
A short intervention on the supervision link, planned with you. The controllers keep running.
Demonstration
Legitimate commands pass, out-of-limit commands are blocked, the record is verified with your teams.
Report
Success criteria agreed at the start and measured. You choose what follows: annual licence or removal.
Reversible: removing OASIS restores the original link. Outside the safety chain: emergency stops and safety functions remain the equipment’s own. Architecture and evidence file are shared under a non-disclosure agreement.
Questions
Questions we are asked.
Will OASIS block a legitimate command?
Only if it is outside the rules you set: a value out of limits, an unknown or unauthorised operator, a missing second approval, whoever sends it. Inside the rules, every command passes. One identified operator is enough for routine adjustments; two different people are needed only for the commands you mark as critical, and two approvals never allow a value outside the limits.
Does it slow production?
On our bench, OASIS adds about 2 milliseconds of its own per command, record included, measured against a simulated device. The pilot measures it on your line.
What happens if OASIS itself stops?
The equipment keeps running its program; remote changes wait. What operators do in the meantime, and what they do when a command they need is blocked, is agreed with you at scoping and exercised during the pilot.
What is installed on my side?
Nothing in the controller program and nothing in the supervision software. OASIS is placed on the link between them. How your operators are identified is set up with you at scoping; the detail is part of the confidential documentation.
Does OASIS touch the safety chain?
No. It sits outside the safety chain and replaces no safety function; emergency stops and safety functions remain the equipment’s own.
What about the other ways into the controller?
During scoping, every access to the controller is listed. OASIS protects the path through which the supervision writes; the inventory shows which other paths must be closed.
What about AI agents?
OASIS does not try to recognise the attacker. It checks each command against the same limits, whether it comes from a person, a script or an AI agent. The check follows fixed rules; it has no model that could be persuaded or retrained. AI agents that run equipment within limits set by people are on our roadmap.
Is it proven outside your lab?
Not yet. Our figures come from our own bench, on one commercial controller. Reproduction by a third party is planned for Q4 2026 to Q1 2027, and each pilot repeats the tests on the customer’s line.
See OASIS running on a real controller.
A live session with our team: your scenario, your limits, and what passes or is blocked. Send us your controller and supervision models and we confirm compatibility, with no commitment.