Determinism vs Non-Determinism: Securing AI Applications and AI Agents

A10 Networks

Jul 30, 2026

In this session, Jamison Utter and Arjoyita Roy from A10 Networks discuss how artificial intelligence fundamentally shifts traditional cybersecurity paradigms. Learn why classic security models fall short when defending probabilistic AI systems and how to re-architect security controls for modern AI workloads.

Key Discussion Points

Predictable vs. Probabilistic: How traditional software relies on deterministic rules
($f(x) = y$), whereas AI systems operate on probabilistic predictions ($P(y \mid x, \theta)$).
Application vs. Intelligence Layer: Securing the deterministic code-based stack (APIs, authentication) vs. defending the non-deterministic reasoning engine.
New AI Threat Vectors: Understanding risks like Prompt Injection, Goal Drift, Tool Misuse, and Memory Poisoning.
5 Core Steps to Securing AI Agents:
a. Constrain Action Surfaces: Enforce strict schemas and capability-based access control.
b. Separate Planning & Execution: Avoid giving a single agent total authority across the entire workflow.
c. Require Structured Reasoning: Validate step-by-step execution plans rather than exposing raw chain-of-thought.
d. Add Oversight Models: Deploy critic and policy-checker LLMs to maintain statistical governance.
e. Shift from Observability to Explainability: Trace log sequences to explain decision branches when anomalous behavior occurs.

Core Takeaway: You don't secure the model, you secure the system that contains the model. Security must evolve from static rules to statistical governance.

Timestamps

00:00 - Introduction: Securing AI Workloads

01:48 - Predictable vs. Probabilistic Software Models

04:38 - Why Zero Trust Needs Re-Evaluating for AI

07:00 - Securing the Deterministic Application Layer

11:00 - Securing the Non-Deterministic Intelligence Layer

15:34 - Unique Threats to Non-Deterministic Systems

19:48 - 5 Best Practices for AI Agent Security

25:15 - Key Takeaways & Conclusion

Learn more about A10's AI Firewall: https://bit.ly/4tIdke1

Visit A10Networks.com for more insights on enterprise AI security and infrastructure.

#cybersecurity #aisecurity #aiagents #llmsecurity #a10networks #appsec #zerotrust #infosec #artificialintelligence #tech #devsecops