Adoption of agentic AI coding and cowork agents such as Cursor, Claude Code, Claude CoWork and Codex is skyrocketing. Security teams have been scrambling to meet the challenges this of the new attack surface created by the complex fabric of interacting AI agents, skills, MCPs, hooks, connectors and other components on user laptops and workstations.

We've already seen how tools such as OpenClaw can break trust boundaries and create backdoors into endpoints, but even commercial tools with better security controls and permissions can be manipulated once extended via external tools using MCPs, hooks and skills.

In this talk we will cover:

• What the new ecosystem look like, and why it represents a huge blind spot for security teams
   
• The threat surface it creates and what unvetted and unsupervised agentic AI dev tools and MCPs can do, from prompt injection to privilege escalation and RCE
   
• Why the defaults and internal security controls provided by vendors such as Cursor are not sufficient
   
• Which best practices and free tools can be used for discovery, visiblity and hardening of this new stack