← Home
TWIML · June 16, 2026 · 56m

Why AI Agents Break the GenAI Security Model

Devvret Rishi
Show Notes

Dev Rishi, GM of AI at Rubrik, explains why traditional enterprise security approaches fail when AI agents move beyond answering questions to taking autonomous action across tools and systems. The episode explores how agent capabilities — planning, tool calling, code generation, and workflow automation — create new security vulnerabilities that static guardrails and human approval can't contain, and what defense-in-depth approaches enterprises need instead.

This summary was generated from show notes and public descriptions, not from a full transcript review. Details may contain inaccuracies.

Curious

•
Runtime Enforcement and Observability Replace Preemptive Rules
Enterprise security needs to shift from writing rules in advance to observing agent behavior at runtime and enforcing policy dynamically — catching harmful patterns as they emerge rather than trying to prevent them preemptively.
ClaudeChatGPTGemini

Novel

•
Defense in Depth: Agent Rewind and Recovery
When agent actions do cause harm, systems need the ability to detect the mistake and roll back — not just prevent it, but recover from it, because perfect prevention is impossible.
ClaudeChatGPTGemini

Highlights

•
Agent Security Breaks Static Guardrail Model
Traditional enterprise security — static guardrails plus human approval — fails when AI agents operate autonomously across tools and systems because the approval bottleneck can't scale to machine speed and agents can route around controls.
ClaudeChatGPTGemini
•
Tool Access Increases Blast Radius Nonlinearly
Giving agents access to multiple tools doesn't increase security risk linearly — each new tool creates combinatorial risk because agents can chain tools together in unexpected ways.
ClaudeChatGPTGemini

Editorial

•
AI May Be Needed to Secure AI
The complexity of securing autonomous agents may exceed human cognitive capacity, requiring AI systems to monitor, evaluate, and enforce policy on other AI agents.
ClaudeChatGPTGemini

Misc

✧The term 'security theater' applied to human-in-the-loop review when agents operate at scale — humans can't keep up with machine speed
✧Agents can route around controls in surprising ways — problem is more subtle than just restricting tool access
✧The core tension: agents are useful because they're autonomous; that autonomy is also what makes them hard to govern
Was this useful?