Security before autonomy

An agent never gets the keys to the business.

We separate conversation, decision and execution. The model proposes; least-privilege services validate and execute; every action leaves evidence.

Discuss security
AIPITER / SecurityIllustrative journey
4checks before taking action
  1. 01Company and identityIsolated context
  2. 02Tool permissionMinimum scope
  3. 03Business ruleDeterministic validation
  4. 04ConfirmationRecorded evidence
01 /

Business isolation

Separation lives in the database and each service identity, not in an optional model instruction.

  • Row-level policies
  • Non-interactive service roles
  • Secrets kept out of code and models
02 /

Bounded tools

Every tool declares what it can read or change. Sensitive operations require confirmation, authorisation and validation.

  • Capability-based permission
  • Validated inputs
  • No direct LLM database access
03 /

Complete traceability

We retain who, what, when and the outcome of relevant operations so they can be investigated, explained and recovered.

  • Auditable events
  • Cost per operation
  • Alerts and human handoff
04 /

European privacy

We design to minimise data, retention and exposure, using European infrastructure where the provider and channel allow it.

  • Verifiable deletion
  • Customer export
  • Documented providers

Frequently asked questions

Control you can understand.

Can the AI model modify the database directly?

The architecture separates the model from execution. The model proposes using a tool; services with restricted permissions check identity, inputs and rules before acting. Instructions written inside a message must not be able to expand those permissions.

How are different businesses’ data kept separate?

Isolation is designed into service identity, permissions and data access, not just a prompt. Each query and action must remain within the business’s authorised context. The specific configuration and its tests are part of implementation review.

What is recorded about an operation?

The traceability design covers who requested the action, which tool was involved, when it happened and the outcome. The goal is to explain changes or investigate failures without retaining data indiscriminately. Logging scope and retention must be defined for each implementation.

Does this page constitute a security certification?

No. This page explains design and control principles; it does not claim independent certification or guarantee that incidents cannot occur. Evaluation must consider the actual configuration, providers, access permissions and tests for the implementation being used.

Aipiter

An agent never gets the keys to the business.

Discuss security