AI Agents and Infrastructure: Multi-Modality, World Models and Enterprise Adoption
Danny Kirabo (iShield 360 x SCL/IFIS) on the five levels of AI risk, from prompt injection to memory poisoning, the agentic incidents of 2025-2026, and the checks a Uganda institution can run.
Slides from the conference talk by Danny Kirabo (iShield 360 × SCL/IFIS) on how organisations can adopt AI agents without losing control of them. The central idea: a model becomes an agent when it gains context, memory and tools, and an agent becomes a risk when it gains identity and reach.
The deck walks through five levels, from the model itself to the infrastructure it runs on, multimodal input and world models, and sets out the security consequence of each, including prompt injection, tool poisoning and memory poisoning. It classifies the incidents of 2025 and 2026 against the OWASP agentic taxonomy, summarises the controls governments published this year, and closes with Uganda's own picture and checks your institution can run.
Useful as a briefing document before a board or risk committee discussion of AI deployment.
Who it is for
Boards, CISOs, risk and technology leaders weighing AI agent adoption