September 2026 Summaries

Capgemini’s 2026 report identifies AI as the leading priority in digital sovereignty strategies for 75% of organizations, reflecting growing concerns about resilience, control over data and models, provider concentration, export restrictions, and dependence on external infrastructure. Digital sovereignty has become a board-level issue, with most organizations discussing it at that level, many developing formal strategies, and a majority allocating budgets, but the report argues that sovereignty should mean resilient, controlled interdependence rather than complete technological isolation. It distinguishes sovereign AI from sovereign cloud, noting that in-region or locally controlled infrastructure alone does not ensure control over sensitive data, models, access, processing, or governance. A practical approach is “minimum viable digital sovereignty,” in which organizations identify critical workloads and apply stronger controls, portability, verification, and exit options where risks are highest. The text also describes a model involving Intel confidential computing, Red Hat OpenShift, and Duality’s governed collaboration workflows as an example of how organizations could use third-party or public-cloud infrastructure while restricting operator access to sensitive AI assets.

Trustworthy AI is presented as a lifecycle-based approach to developing and operating systems that are reliable, safe, secure, fair, transparent, explainable, privacy-preserving, and accountable, with these qualities demonstrated through auditable technical and governance controls rather than ethics statements alone. Drawing on frameworks such as NIST’s AI Risk Management Framework, the discussion emphasizes that regulated sectors must document model ownership, testing, decision rationales, monitoring, cybersecurity, and data governance. The EU AI Act is described as making many of these expectations legally binding for high-risk systems, with substantial penalties for violations and phased implementation deadlines extending into 2027 and 2028. Healthcare, finance, and government apply these principles through sector-specific validation, oversight, model-risk management, and data-sovereignty practices, while facing a common challenge of using sensitive data held across institutions. Privacy and security are characterized as foundational because conventional encryption may leave data exposed during computation, and privacy-enhancing technologies such as federated learning, fully homomorphic encryption, and differential privacy can enable collaborative AI training and analysis without exposing raw records.

Membership inference attacks use a model’s responses to determine whether a specific record was included in its training data, potentially exposing sensitive information without accessing the underlying dataset or systems. The risk is especially significant in healthcare, finance, and government because dataset membership can reveal diagnoses, financial status, or involvement in sensitive government activities, while large language models may amplify the threat through memorization and possible reproduction of training content. The source argues that GDPR and HIPAA may treat successful attacks as privacy or breach events, reflecting growing regulatory scrutiny of AI models as possible carriers of personal data. Conventional protections such as encryption, access controls, and network security do not directly prevent these attacks because adversaries can use legitimate model queries, while output restrictions, rate limits, and regularization may reduce but not eliminate exposure. It presents differential privacy as the only defense with a mathematical privacy guarantee, and describes combining it with approaches such as federated learning, fully homomorphic encryption, and confidential computing to support privacy-preserving AI development across regulated organizations.