Your rules
Define which data may be used, which destinations are permitted and which actions must remain blocked.
IN DEVELOPMENT · AI-IDEENSCHMIEDE
AI acts.
You decide.
AI protection against AI attacks
We are developing AI-Certain as an independent control layer between AI assistants and the data, applications and APIs they want to access.
This interface is a demo. Technical protection will depend on the security core and verified integrations.
ACTION CHECKConcept example
An assistant requests an actionReview tomorrow’s reservations
AI-CERTAINCheck first.
Check first.
Then execute.
TaskPermissionsData & destination
AllowWithin scope
HoldApproval needed
BlockBoundary breached
Enforced boundaries outside the AI model.
01 / OUR APPROACH
Responsibility
stays with people.
An AI assistant can analyse information and propose actions. AI-Certain is designed to check whether a requested action fits the task, which permissions apply and what data would leave the system.
Our security approach combines technical boundaries with decisions you can understand. Critical actions are held or blocked. Approval stays with the user wherever their rules require it.
Our principleAn AI must not expand its own permissions or disable its own safeguards.
02 / PLANNED FEATURES
A framework for
your AI assistants.
These building blocks guide our development.
Tasks & permissions
Give each assistant a specific task with appropriate permissions, a time limit and a cost limit.
Multiple AI models
Choose models through supported integrations and divide tasks between assistants with separate roles.
Specific approvals
Review the exact action and its effects. Each approval is intended to apply once to that specific action.
Emergency stop
Stop one or all assistants using a visible button or your personal stop phrase through your authenticated control channel.
A clear activity log
Track what was requested, allowed, held or blocked, with clear reasons.
03 / HOW IT IS DESIGNED TO WORK
One task.
Three clear boundaries.
Example task: “Review tomorrow’s reservations.”
Choose an action to see the intended decision.
ALLOWED
Reading is part of the task.
The assistant may read the required reservations. Access remains limited to the defined scope.
Rule: reading the required data is allowed
CONCEPT EXAMPLE · NO REAL SYSTEM ACTION
04 / OUR WORK
Building control
step by step.
AI-Certain is being developed at AI-Ideenschmiede. We combine human experience with AI as a tool and review the results of our work.
Our goal is an understandable security solution with technically enforced rules. A second AI can provide additional insights. Binding access limits are intended to be enforced by the independent security core.
More about AI-Ideenschmiede- DEVELOPED
Concept & interface
Rules, permissions, models, approvals and emergency stop controls have been developed as a concept and interactive demo.
- NEXT STEP
Security core & first integration
Check actions on the server and execute them under control. Permissions and stop controls must be enforced independently of the assistant.
- PLANNED
Test, measure, improve
Test permitted and prohibited actions, manipulation attempts and failures. Use those results to build further integrations.
05 / ONE CORE, SUITABLE INTEGRATIONS
AI-Certain and AILENDO Guard.
STANDALONE SOLUTION · PLANNED
AI-Certain
The shared security core and management interface for connected AI assistants and systems.
INTEGRATION · PLANNED
AILENDO Guard
The same approach, with rules and interfaces for AI features within AILENDO. The user stays in control.
06 / FREQUENTLY ASKED QUESTIONS
What you should know today.
Is AI-Certain already a finished security product?
We are still in development. The integrated interface demonstrates the controls using sample data. It is not connected to an AI model or production system and does not block real attacks.
What does “AI protection against AI attacks” mean?
We focus on controlled AI actions: manipulated content must not give an assistant extra permissions. AI-Certain is designed to enforce the user’s boundaries independently of the model. This does not require identifying an attack as “AI-generated”.
Can I use several AI models at the same time?
That is planned. Assistants can be assigned different models and roles. Communication between them is intended to follow the same data and permission rules. Supported models will be determined through verified integrations.
Can the emergency stop undo everything?
The planned stop blocks new actions. Cancelling an action already in progress depends on the relevant interface. Data already sent and actions already completed cannot be reversed by the stop.
Which systems will AI-Certain protect?
Protection is planned for actions routed through controlled integrations. Each integration requires its own technical verification. Websites, servers and user accounts also require standard security measures against general attacks.
When can I buy AI-Certain?
A release date and price have not yet been set. Distribution is planned through Ailorium. The demo is free to explore.
07 / INTERACTIVE DEMO
Try the controls.
Choose models, draft rules and tasks, review an approval and test the emergency stop. Everything runs locally with sample data. Settings are temporary and do not enforce real security rules.
Control for your AI assistants
Demo: monitoring active
Workspace · AI-Ideenschmiede
Your AI under control.
Interactive concept · Sample data · No connected systemsAssigned assistants
2
Pending approvals
1
Blocked actions
1
Assistants & models
AILENDO assistantActive
Model A · Review reservations
Reading allowed · Changes require approval
Review assistantActive
Model B · Review drafts
Reading allowed · No execution
Your protection rules
01
Customer data stays in the system
Export to external recipients is blocked.
02
Changes need your approval
Reservations and messages.
03
The user controls permissions
Assistants cannot expand their permissions.
Approval required
Change pickup time
AILENDO assistant · Reservation R-1042
5 October 2026
10:00 → 11:00
Customer details remain unchanged. No data is sent externally.
Blocked by policy
External customer data export
Recipient outside permitted destinations. This action is outside the task’s scope.
Model assignment
Switching providers requires appropriate permission to share data. In a future implementation, keys will be stored securely on the server.
Rules for this workspace
In the planned product, written rules will be translated into verifiable settings and confirmed by you before activation.
Create a task
Reservation R-1042 · Change pickup time
5 October 2026 · 10:00 → 11:00
Assistant: AILENDO assistant · Model A
Affected data: pickup time for this reservation
External recipient: none
Approval applies once to this exact change.
Decision log
Blocked · Customer data export
Rule: recipient not permitted · No transfer
Rule: recipient not permitted · No transfer
Stopped · Change pickup time
Rule: changes require user approval
Rule: changes require user approval
Allowed · Read reservations
Within task scope · Permissions checked
Within task scope · Permissions checked
Your emergency stop
In the planned product, the stop phrase will only be accepted through your authenticated control channel. Documents and third-party messages will not be able to trigger it. This demo does not implement authentication.
PEOPLE + AI + CLEAR BOUNDARIES
Take a closer look
at our work.
Explore the interface and see how rules, approvals and stop controls are designed to work together.