From MCP to Agentic AI: Can Agents Uphold Ethics? | Prof Siti Hafizah | IEEE CS Malaysia

📅 Recorded live on 27 October 2026 · IEEE Computer Society Malaysia Chapter Webinar Series AI agents are moving from answering questions to acting on our behalf, connecting to tools and data through protocols such as the Model Context Protocol (MCP). This webinar traces the path from MCP to agentic AI and asks whether autonomous agents can be trusted to uphold ethics: who is accountable when an agent exceeds its authority, what controls keep its actions within clear boundaries, and what evidence we should require before delegating the next task. What you will learn: • The speaker opened with news cases from September 2026 in which AI agents allegedly collected data from Australian and Indian government websites without permission, raising the question of who answers for an agent's actions. • Agentic AI starts from a user's goal, plans, uses tools, checks its own results and may change databases and other resources, while the Model Context Protocol (MCP) connects models to tools and data in one shared way. • An agent asked to send a report to a committee may read private folders on the way, so the speaker asked whether every action can be reconstructed and reversed, and pointed to the OWASP agentic security initiative's ten risks. • She proposed two controls: consent that is enforced in the system, with approvals bound to the exact operation, and treating the whole action sequence, including retries and partial failures, as the unit of assurance. • Execution evidence should be stored where the agent cannot edit or delete it, summarised as OCAVR: observe and constrain, audit and verify, recover. • Four case studies (music emotion tagging, museum review analysis, coding agents that weaken failing tests or edit audit logs, and a continually learning intrusion detector) each showed where an ethics checkpoint belongs. • In the Q&A the speaker argued that governments must regulate in line with national needs and that computer scientists should work with social science and law experts, while the moderator noted that an AI governance bill is expected to be tabled next year. Chapters 0:00 Welcome and housekeeping 1:52 Opening: who answers for an agent's actions? 3:52 What is agentic AI, and how it differs from ordinary AI 6:52 MCP: the Model Context Protocol 8:52 Hidden risks: private data and reconstructing actions 9:52 OWASP agentic security risks 11:52 Stress testing agents: Anthropic's fictional scenarios 14:52 Two controls: enforced consent and the action sequence 17:52 Audit evidence the agent cannot edit 18:52 OCAVR: observe, audit, verify, recover 19:52 Case studies 1 and 2: music emotion tags and museum reviews 26:52 Case study 3: coding agents that change what counts as success 29:52 Case study 4: an intrusion detector that keeps learning 33:52 Human oversight and approval fatigue 34:52 Does trust need evidence? Five research issues 36:24 Q&A: who is accountable for agents? 38:48 Government, regulation and the role of AI companies 41:32 Chat questions: reliable AI code and research opportunities 44:44 Agentic AI in healthcare and other critical systems 47:36 Is Malaysia ready? Budget 2027 and the AI governance bill 52:24 Closing remarks 🎤 Speaker: Professor Dr Siti Hafizah Ab Hamid, Deputy Dean, Universiti Malaya, and core member of the Centre of Research for Cyber Security & Network (CSNET) 🎙️ Moderator: Ts. Dr. Muhammad Faiz Bin Mohd Zaki, Director, Data and Information Management Centre, Universiti Malaya 📅 27 October 2026 · IEEE Computer Society Malaysia Chapter Webinar Series and Professional Workshops, co-sponsored by the Centre of Research for Cyber Security & Network (CSNET), Universiti Malaya 🌐 CSNET: https://csnet.um.edu.my #IEEE #IEEEComputerSociety #AgenticAI #MCP #AIEthics #ResponsibleAI #CSNET #UniversitiMalaya