Beyond Guardrails: Build a Security Harness for Your AI
Ashish Bhadouria
About this talk
AI agents are no longer just answering questions. They can write and execute code, call APIs, access internal systems, use credentials, and make decisions autonomously. That changes the security problem.
The recent Hugging Face incident showed what can happen when an autonomous AI agent gets too much freedom and the surrounding controls aren't strong enough. It raises an uncomfortable question: What happens when our AI does something we didn't expect?
This talk introduces the AI Security Harness, a security layer built around AI that assumes the model will sometimes make the wrong decision. Like a harness around a powerful machine, it doesn't limit what the AI can do; it makes sure it can't go somewhere it shouldn't.
We'll explore practical controls including agent identity, least privilege, tool-call authorization, sandboxing, runtime monitoring, anomaly detection, data-loss prevention, and auditability. We'll also look at continuously attacking and testing the harness itself.
Speaker
Ashish Bhadouria is a cybersecurity and AI security leader with extensive experience building security programs, securing cloud and software platforms, and navigating security, governance and compliance at scale.
His work sits at the intersection of AI security, application security, cloud security and governance, with a focus on making emerging AI systems secure enough to operate in the real world. He has worked across engineering and security leadership roles, helping organizations build secure development practices, protect modern cloud infrastructure, and turn security requirements into practical engineering controls.
Ashish focuses on the security challenges emerging from AI agents, autonomous systems and AI-enabled software development particularly how organizations can move beyond traditional guardrails and build security controls that operate around AI systems at runtime.