Executive Summary
Why Secure AI is a priority
The challengeis no longer simply to protect AI, but to maintain control over its decisions and actions.
The DGS Perspective
Cybersecurity and AI: a single risk perimeter
AI for Cybersecurity
Use AI to enhance prevention, detection, correlation, investigation and incident response capabilities.
Cybersecurity for AI
The new AI threat landscape
AI is profoundly reshaping the threat landscape. On the one hand, it makes established techniques such as phishing, impersonation, social engineering and automated reconnaissance more efficient and scalable. On the other, it introduces new attack surfaces related to models, datasets, RAG systems, knowledge bases and autonomous agents.
Organizations therefore face a dual challenge: managing the acceleration of traditional risks while addressing threats specifically designed to target AI systems. Prompt injection, data poisoning, model extraction, manipulation of retrieval systems and the misuse of agents are not theoretical scenarios, but vulnerabilities that can lead to data compromise, unintended actions and loss of control over business processes.
The priority for management is to identify which threats create immediate exposure, which require monitoring capabilities and which must be governed through structural controls throughout the AI lifecycle.
From output control to action control
With generative systems, the main risk factor concerned the quality or reliability of the responses produced. The evolution towards RAG architectures and agentic systems shifts the focus to the actions that artificial intelligence can perform through digital identities, connectors and enterprise applications. An agent can query knowledge bases, access systems, retrieve data, generate workflows and interact with operational processes. Control therefore cannot be limited to the model itself: it must extend to privileges, authorizations, non-human identities, decision traceability and action reversibility. The guiding principle of Secure AI is simple: no level of autonomy should be introduced without adequate observability, explicit accountability, approval mechanisms and the possibility of human intervention.
Secure AI Lifecycle: designing control throughout the entire lifecycle
AI security cannot be added at the end of an already completed project. It must accompany every stage of the journey, from use-case definition through operational management and system decommissioning. A Secure AI approach integrates risk assessment, data classification, dependency control, model validation, adversarial testing, continuous monitoring and incident management within a single governance framework. Each stage must produce verifiable evidence, assign clear accountability and define objective release criteria. The goal is not to slow down AI adoption, but to make it sustainable, measurable and controllable over time.
What Is DGS's Approach to Secure AI?
DGS's experience in integrating cybersecurity, AI & Data, cloud, governance and managed services enables us to address AI security as an operational challenge rather than an exclusively technological one. The DGS vision is based on a fundamental principle: using AI to strengthen defense capabilities while protecting the AI systems adopted by the organization. These two needs share data, controls, processes, telemetry and accountability. For this reason, governance must be unified and focused on continuous risk measurement. Secure AI thus becomes a discipline that integrates assessment, governance, security engineering, monitoring and compliance within a single control model.
DGS capabilities for Secure AI
The goal is to build demonstrable and progressively mature capabilities, in which security, compliance and innovation evolve in a coordinated way and remain aligned with the organization's risk profile.
Il contesto operativo
Roadmap for sustainable Secure AI
Cosa significa nel concreto
- Build visibility and control
AI adoption requires a progressive journey that allows organizations to increase the level of autonomy only when the corresponding controls are fully operational. The goal is not to slow down innovation, but to build a sustainable capability over time, balancing speed, security, governance and accountability. Secure AI maturity is therefore measured by an organization's ability to understand, monitor and govern the behavior of AI systems throughout their lifecycle.
Once the perimeter has been defined, responsibilities and operating rules must be established:
- Definition of policies and usage criteria.
- Assignment of decision-making responsibilities.
- Integration across Cybersecurity, AI, Risk & Compliance.
Introduction of approval and oversight processes.
3. Embed security by designSecurity must accompany the design, deployment and operation of AI systems:
- Risk assessment and use-case classification.
- Threat modeling and control validation.
- Testing and simulation of adversarial scenarios.
Continuous monitoring of data, models and agents.
4. Measure and demonstrate control
Trust requires verifiable evidence and explicit residual risk:
- Logging and audit trails.
- KPIs and KRIs dedicated to AI systems.
- Monitoring of exceptions and overrides.
- Continuous production of evidence for audit, compliance and management.5. Scale autonomy in a controlled way
Autonomy can progressively increase only when appropriate control mechanisms are in place:
- Human oversight of critical processes.
- Management of privileges and non-human identities.
- Rollback and kill-switch mechanisms.
- Periodic review of risks, controls and autonomy levels.
Conclusions
Secure AI is a necessary step in turning experimentation and innovation into reliable, measurable and sustainable capabilities. It means integrating security, governance, compliance and operations within a single model capable of delivering control, accountability and verifiable evidence throughout the entire AI lifecycle.
True maturity is not measured by the number of models adopted or agents deployed, but by the ability to grant autonomy progressively while always retaining the ability to observe, understand and govern the consequences of their actions.
L'adozione di un'architettura multisito centralizzata permette di gestire diversi touchpoint digitali da un unico backend. Questa flessibilità consente al marketing di lanciare rapidamente nuove offerte commerciali e mini-siti dedicati, mantenendo una forte coerenza visiva e di brand, e semplificando la governance complessiva dei contenuti.
Expert voices
H3
