Lightning Talk: Your Most Privileged User Isn't Human: The Agent Security Blinds...
summary
Kubernetes clusters end up giving AI agents broad service-account powers so a nonhuman identity makes unauditable changes: audit logs show only a service account, agents use kubectl and cloud CLIs to read secrets and escalate, and agents dynamically invoke multiple tools / MCP servers creating long attack paths (example: a misconfigured GitHub Action led to a supply‑chain compromise). Fixes: inject lost attribution by annotating user context, enforce permission escalation controls with runtime policies, and capture tool chains via OpenTelemetry / GenAI tracing for invisible tool chains.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 1:49AI agents as privileged users
- 3:14Gaps in agent security
- 8:29Solutions for agent security
- 10:16Closing remarks
related talks
27mSecurity by design for agentic AI: How CISOs stay in control of autonomous systems
25mWhat We Learned Securing AI Agents at Scale on Multi-Tenant Kubernetes Clusters
40mIn AI We Trust? Securing the Future, One Agent at a Time
45mYour AI Agent Has Production Access: Now What? (Jack)
09mLightning Talk: Securing K8s Pods from Within: A Runtime Approach
10m⚡Lightning Talk: Securing AI Agents: Applying Service Mesh Patterns To MCP
27mKeeping Your Agents in Check: Layered Security for Agentic Platforms in Production
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust