AI Technologies & Controls

Where governance sets the rules and risk management decides what to worry about, this page is about the hands-on security engineering of AI: designing secure AI architecture, securing the model life cycle, controlling data, embedding privacy and safety, and running the control and monitoring program that keeps it all assured.

Secure AI architecture and design

AI systems should be secure by design and integrated deliberately into enterprise architecture:

• Secure development life cycle (SDL) applied to AI, with change management over models, data, and infrastructure.
• Securing infrastructure as code, so the environments that provision AI are hardened by default.
• Mapping and securing data flows through AI pipelines.
• Approved base models — governing which foundation models may be used, and under what conditions.
• Interconnectivity — securing how AI systems integrate with, and are reachable from, the rest of the enterprise estate.

Securing the model life cycle

Models need security and assurance across selection, training, validation, and deployment:

• Model testing and evaluation — measuring accuracy, reliability, and security behaviour.
• Interconnectivity and linkages — validating how models interact with each other and with systems, since one model's output often becomes another's input.
• Regression and progression control — ensuring changes don't reintroduce prior faults or degrade performance, and promoting models through environments in a controlled way.
• TEVV (Testing, Evaluation, Verification, and Validation) — the structured assurance discipline that underpins trustworthy AI.

Data management controls

Because a model's behaviour is learned from data, the data is part of the attack surface and part of the control surface:

• Data collection and control — governing how data is gathered, with access, integrity, and provenance controls.
• Data poisoning defences — protecting against malicious manipulation of training data.
• Bias and accuracy — detecting and mitigating bias, and maintaining data quality that supports correct behaviour.
• Data positioning requirements — constraints on where and how data is held, including residency.

Privacy, ethical, trust, and safety controls

These controls keep AI lawful, ethical, transparent, and safe to use:

• Explainability and transparency — making decisions interpretable, and disclosing where and how AI is used.
• Privacy controls — honouring data subject rights (including the right to be forgotten), lawful consent, and data minimisation and anonymisation.
• Fairness and ethics — preventing discriminatory outcomes and operationalising ethical principles.
• Human oversight — meaningful human-in-the-loop review, especially over automated decision making.
• Trust and safety — content moderation of AI output, assessment of potential harm, and consideration of environmental impact.

Security controls and monitoring

Finally, controls have to be selected, implemented, and continuously assured:

• Selecting and implementing the right controls, matched to assessed AI risk, and mapping threats to controls so coverage is demonstrable.
• Control self-assessment (CSA) and a managed control life cycle, so controls stay effective as systems and threats evolve.
• Continuous monitoring with defined security monitoring metrics, and KPIs/KRIs that measure both control effectiveness and residual risk.
• Security awareness training that builds AI-specific awareness across the workforce.

Related pages

• AI Governance & Program Management
• AI Risk Management
• Secure System Design · Application Security Testing
• Data Security · Analytics & Detection

References

• NIST AI RMF — Generative AI Profile
• OWASP Machine Learning Security Top 10
• ISO/IEC 42001 — AI management systems