Protection and resilience
The S.H.I.E.L.D. methodology helps you plan the protection of the systems AI works in: from an inventory of assets to incident detection and response. It also covers operational technology (OT) environments, such as production control systems.
- Intent
- Input control (applies)
- Authorization (applies)
- AI (applies)
- Result validation
- Human
- Evidence (applies)
Highlighted: the points of the flow that this description refers to. Orange marker: a human decision at this point.
What problem the methodology addresses
AI introduces new points of contact, that is, models, connectors to documents, tools and AI digital workers that perform actions. Industrial companies also add contact with OT systems. Each of these points must be covered by the protection rules that apply to the rest of the infrastructure.
New ways for data to leave
Content passed to models and tools can leave the organization if the path is not controlled.
New entry points
Documents processed by AI may contain instructions that try to change how it behaves.
Dependence on services
You need to know what happens to a process when a model, a provider or a connection stops working.
What the procedure looks like
The six S.H.I.E.L.D. steps structure protection from the list of assets to incident response. The methodology complements the organization’s security rules but does not replace them. The letters of the acronym come from the English names of the steps.
Step 1 · S (Security Baseline)
Inventory and baseline
List the assets AI runs on or connects to: servers, models, connectors to data sources, service accounts and, in industry, also OT devices and systems. Set minimum security requirements for each group.
Step 2 · H (Hardening Assessment)
Hardening
Check network segmentation, firewall rules, permissions and the configuration of the systems AI runs on. Limit the permissions of accounts and services to what is necessary.
Step 3 · I (Integration Points)
Integration points
Describe where AI connects to IT and OT systems and to external services. For each connection, decide what data may flow through it and who approves it.
Step 4 · E (Edge & Endpoint)
Edge devices and endpoints
Include edge devices, industrial sensors and workstations. Blocking data from being sent from private accounts, browsers or other applications requires controls on the organization’s devices and network.
Step 5 · L (Legacy Evaluation)
Legacy systems
Assess the systems that cannot be updated quickly. Set additional safeguards and restrictions on AI access for them.
Step 6 · D (Detection & Response)
Detection and response
Decide which events are monitored, who responds to an incident and how to restore operation. Rehearse a scenario in which a model, a provider or a connection fails.
Which steps of the method AiTrust supports
A methodology describes how an organization sets the rules for working with AI; AiTrust provides technical support for selected steps, to an extent that depends on the deployment configuration.
On-premise deployment
- By default in AiTrust
- In the local option, the AiTrust application, the local model, and the data and documents run in the organization’s infrastructure.
- Your organization decides
- The Polski Instytut AI team installs and configures it together with the organization’s IT department. We agree the responsibility for the infrastructure, the backups and the way updates are installed before the launch.
Related steps
Inventory and baseline (S), Hardening (H)
Controlled query path
- By default in AiTrust
- Queries in configured paths go through AiTrust.
- Your organization decides
- The organization controls traffic outside the configured paths on its devices and network.
Related steps
Integration points (I), Edge devices and endpoints (E)
Data sanitization before the model
- By default in AiTrust
- The model receives the context after sanitization of designated data categories.
- Your organization decides
- The organization sets the rules and the scope of detection. Masking on its own does not mean irreversible anonymization.
Related steps
Integration points (I)
Local models and working without the Internet
- By default in AiTrust
- AiTrust can work in the local option, in which the application, the local model, and the data and documents run in the organization’s infrastructure. Local models connect through Ollama and vLLM.
- Your organization decides
- Working offline depends on the configuration and the process’s dependencies, and it requires local models and all the services needed.
Related steps
Integration points (I), Detection and response (D)
Suspending or revoking an AI digital worker
- By default in AiTrust
- An AI digital worker without an active agent passport will not use tools.
- Your organization decides
- If needed, the organization suspends or revokes the agent passport based on a review of the recorded events.
Related steps
Hardening (H), Detection and response (D)
Evidence trail
- By default in AiTrust
- Operations in the controlled path leave an evidence trail that can support incident analysis.
- Your organization decides
- The organization sets the evidence retention period.
Related steps
Detection and response (D)
How this step looks in the AiTrust application
The output firewall checks responses according to the selected mode: strict, standard or custom.
“Weakening the sanitization policy requires the consent of a second, different person; tightening it takes effect immediately.”
What the available example shows
We do not publish a separate screenshot for this methodology. The protection of networks, devices and systems depends largely on the organization’s own safeguards, not only on the AI platform.
We show an example of data sanitization before sending to the model on the Human oversight of AI page.
Related answers on the platform and compliance
Answers on the platform page and on the AI Governance and compliance page
- On-premise deployment
What is needed to run AI in the organization’s infrastructure.
- Working without Internet access
What determines whether a process can work without an Internet connection.
- Sending data outside the company
Where controls on devices and the network are needed.
- Audit
Information on audit in the security and compliance area.
Other methodologies
- AI operating boundariesB.O.R.D.E.R.
- Human oversight of AIH.A.D.A.
- Readiness assessment methodologyI.A.G.R.A.
Let’s discuss AI protection in your infrastructure
In the demo, we will show the controlled query path in a demonstration configuration and discuss which aspects of protection depend on your network, your devices and your deployment option.