
Ep # 16 - Policy Enforcement for AI Agents
Notas del episodio
Policy Enforcement for AI Agents: Why Human Approvals Can’t Keep UpAI agents are requesting access at machine speed, chaining MCP tool calls across ten or more applications and spawning sub-agents. Access policies built for humans waiting on approval tickets can’t keep pace.Sameer Mehta, Chief Product Officer at Opal Security and former product leader at Veza, joins Lalit Choda (Mr. NHI) on the Non-Human and AI Identity Podcast. He explains why policy enforcement for AI agents needs a new model. That model treats agents as mission-driven software with no conscience, combines fixed rules with AI-driven decisions, and stops granting standing access based on “vibes.”The conversation covers why agents are already reaching production without proper security guardrails, and how to keep policy as code predictable when the agents it governs are non-deterministic. It also covers how agent manifests expose off-purpose agents, and why access decisions only scale when an AI agent approves, denies or escalates them.What you’ll learn:Why identity security doesn’t need to be thrown away for agentic AI: the problems are the same, but the scale and attack surface have changedHow standing access piles up in production through judgment-call approvals, and why it goes unnoticedWhy AI agent security and NHI security share ownership challenges but are not the same thingWhy one-size-fits-all access policy breaks for agents, and how to set policy by identity type, resource sensitivity and agent lifespanHow to combine fixed rules and custom instructions with AI-driven access decisions so outcomes stay predictableHow agent manifests reveal off-purpose agents that stray beyond their declared MCP tool call chainHow to evolve policy from access-decision data: move routine requests to birthright access or auto-approval, and keep humans in the loop for privileged resources and break-glass accountsKey Moments:Introduction & Meet Sameer [00:00–03:15]Mr. NHI’s Human Identity in the Hot Seat [03:15–19:36]What Security Teams Are Asking About AI Agents [19:36–24:16]Is Agent Risk Exposing the Wider NHI Problem? [24:16–25:52]Why Human-Era Access Policy Breaks for Agents [25:52–28:44]Building a Policy Model for Agents: Paladin & Policy as Code [28:44–32:32]Policy as Code vs Non-Deterministic Agents [32:32–36:43]Real-World Example: Custom Instructions & Human-in-the-Loop [36:43–40:00]From Humans to Agents: How Paladin & Opal 0 Evolved [40:00–45:01]Agent Manifests & Detecting Off-Purpose Agents [45:01–47:12]Upfront Policy vs Learned Policy [47:12–50:30]Advice for Security Leaders: Start, Measure, Iterate [50:30–51:50]Learn More About Opal & Closing [51:50–53:31]📚 NHI Knowledge Centre: nhimg.org🔗 Opal Security: opal.dev#NHI #NonHumanIdentity #AIAgents #AgenticAI #IdentitySecurity #MrNHI #DigitalIdentity #IAM #PolicyAsCode #PolicyEnforcement #AccessControl #IdentityGovernance #IGA #LeastPrivilege #JustInTimeAccess #MCP #AgenticSecurity #Cybersecurity #MachineIdentity #CISO #AccessManagement #PAM #ServiceAccounts #AgentSecurity #ArtificialIntelligence