Securing the multi-cloud control plane in the age of AI agents

While organisations are making AI adoption a priority, few are addressing the structural collapse of the traditional security perimeter. As organisations evolve from using simple LLM chat interfaces to autonomous agents capable of writing code and managing cloud infrastructure, the attack surface has undergone a phase shift: we have moved beyond prompt injection to the full-scale compromise of the enterprise environment and its core identities.

In this session, Mandiant practitioners Rupa and Jon provide an evidentiary post-mortem of AI security derived from the 2025-2026 frontlines. We argue that the traditional definition of Tier 0 which was once limited to Domain Controllers and Root CAs; must now expand to include the modern administrative control plane: cloud identity platforms, CI/CD pipelines, IaC, and, crucially, the AI agents that orchestrate them.

We begin by categorizing the enterprise AI landscape into four distinct archetypes namely Consumers, Integrators, Trainers, and Builders. This will help reveal how attackers exploit specific implementation flaws to gain a foothold. Through technical case studies and live demonstrations (evolving research previously showcased at RSAC), we will demonstrate how threat actors hijack high-privilege "Agentic" identities to move laterally into the management tier of multi-cloud environments.

The centerpiece of this session is the introduction of the "Tiered AI Security Model." This is a holistic, validated framework designed to harden the four pillars of the modern AI stack: the data, the model, the underlying infrastructure, and the agentic identities themselves.

Attendees will leave with a technical blueprint to ensure AI adoption remains a business accelerator rather than a risk as an automated backdoor into the enterprise. It is time to redefine Tier 0 before the attackers do it for us.