On May 6 2026 Prof. Mentzas talked about the trustworthiness challenges of Agentic AI systems at the closing plenary panel of the “Data Week 2026” conference in Oslo, organized by BDVA – Big Data Value Association—the premier European gathering for Big Data and Data-Driven AI research and innovation.

The session explored cutting-edge advancements in AI.
Check out the full program details here: https://data-week.eu

The main points of his presentation were as follows:

• The shift from Generative AI to Agentic AI is not incremental, it is structural. While Generative AI systems generate content on demand, Agentic AI systems perceive, plan, and act autonomously over extended task horizons. This moves the governance challenge from regulating outputs to constraining consequences.
• Agentic AI systems introduce new classes of risks that current frameworks were not designed for. Goal misalignment, cascading errors across multi-agent pipelines, indirect prompt injection, and persistent memory poisoning are threats that existing governance approaches do not yet address adequately.
• There are many new research challenges that we must address. They include runtime goal alignment, balancing autonomy and control, dynamic (not just pre-deployment) auditing, responsibility attribution and behavioural benchmarks that go beyond task accuracy.