Securing AI Beyond the Filter Layer | A10 Networks
Securing AI Beyond the Filter Layer | A10 Networks
In this second session of our AI security series, Jamison Utter and Arjoyita Roy from A10 Networks dive into the limitations of traditional guardrails and explore what it takes to secure agentic AI. Learn why basic prompt filters fall short, how external guardrails fit into your security posture, and how to govern autonomous AI agents across the entire reasoning loop effectively.
Key Takeaways & Key Highlights
▫️ The Guardrails Illusion: Why standard boundary filters are not enough to secure complex AI behaviors and reasoning paths.
▫️ Built-in vs. External Guardrails: The trade-offs between default provider toggles and customized external security rules for precision and calibration.
▫️ Securing Agentic AI: Moving beyond message-level security to protect multi-step AI agents, tool calls, and reasoning workflows.
▫️ Context Jumping & Escalation: Understanding the security risks when AI agents interact with external tools and pass authorization across contexts.
▫️ Control Loop Framework: Implementing scoped permissions, discovery, kill switches, and planner/executor separation for full agent governance.
Timestamps
00:02 - Introduction & Overview of AI Security Beyond Filters
00:53 - The Guardrails Illusion: Boundary Checking vs. Behavior
02:18 - Why Built-in Guardrails Feel Like Enough (and Why They Fall Short)
04:27 - The Second Wall: Introducing External Guardrails
06:36 - Moving Beyond Messages: The Challenge of Agentic AI
10:13 - Understanding Context Jumping and Privilege Escalation
14:32 - The Contradiction: Where Guardrails Still Fall Short
16:51 - What It Takes to Control an Agent: Scoped Permissions, Discovery & Kill Switches
19:02 - The Takeaway: Agent Protection Framework & Conclusion
🌐 Learn more about securing your network infrastructure and modern AI workloads at: https://bit.ly/4tIdke1
#a10networks #aisecurity #agenticai #cybersecurity #llmsecurity #guardrails #aigovernance #zerotrust #cloudsecurity #machinelearning