Conference Presentation, Keynote
It Takes AI to Break AI: Securing the Agentic Attack Surface | Alice | RAISE Summit 2026
- Enterprises face escalating security and safety challenges driven by the shift to agentic AI capabilities, which enable bad actors to scale fraudulent operations, deepfake creation, and human trafficking activities without human intervention.
- A significant portion of enterprises is expected to fail in deploying AI agents and JNI-based applications to production due to persistent difficulties in securing these tools, testing non-deterministic behaviors, and comprehending associated risks.
- The attack surface will expand as autonomous agents gain the ability to execute actions, make web calls, and reason beyond static chatbots, introducing new threat vectors such as reasoning manipulation and tool calling attacks.
- Organizational ownership of AI security remains ambiguous, with end-to-end responsibility fragmented across security, privacy, product, engineering, and legal teams, hindering effective risk management.
- Traditional pre-deployment controls including model evaluation, red teaming, pen testing, guardrails, automated scanning, and keyword detection are expected to prove insufficient against the non-deterministic nature of AI models, where minor prompt variations can trigger policy violations.
- Alice plans to continue deploying agentic red teaming tools designed to continuously interact with systems, simulate real user behaviors, and evolve via feedback to identify and exploit new bypass techniques for security and safety measures.
- Testing strategies are expected to shift focus from the underlying system to the dynamic behavior of agents, requiring coverage of regulations, compliance, and specific policies throughout the entire development lifecycle and production environment.
- The rapid expansion of the technology surface is projected to exacerbate risks and maintain uncertainty regarding organizational governance as AI models integrate into broader operational workflows.