What's next: agent teams and multi-day tasks

Jim Love discusses how rapid adoption of agentic AI is repeating the industry pattern of shipping technology without security, citing issues like vulnerabilities in Anthropic's MCP and insecure open-source agent tools. He interviews Ido Shlomo, co-founder and CTO of Token Security, who argues AI agents are fundamentally hard to secure because they are non-deterministic, have infinite input/output space, and often require broad permissions to be useful.

Cybersecurity Today would like to thank Meter for their support in bringing you this podcast. Meter delivers a complete networking stack, wired, wireless and cellular in one integrated solution that's built for performance and scale. You can find them at Meter.com/cst

Shlomo proposes focusing security on access, identity, attribution, least privilege, and auditability rather than trying to filter prompts and outputs, and describes Token's "intent-based permission management" approach that maps agents and sub-agents as non-human identities tied to their purpose and allowed actions. The conversation covers real-world risks such as developer tools like Claude Code running with extensive access, widespread over-provisioning of admin permissions and API keys, exposure of unencrypted local token files, and misconfigurations that leak data publicly. Shlomo recommends organizations build governance processes for agents—discovery/inventory, boundary setting, continuous monitoring, and secure decommissioning—and says AI is needed to help police AI. He also highlights emerging trends like agent teams and multi-day autonomous tasks, and notes Token Security is a top-10 finalist in the RSA Innovation Sandbox 2026, planning to present an intent-and-access-focused security model for AI agents.

00:00 Sponsor: Meter's integrated networking stack 00:19 Why agentic AI security is breaking (MCP & open-source chaos) 02:53 Meet Token Security: practical guardrails for AI agents 04:57 Why you can't just ban agents at work (shadow AI reality) 06:24 Tel Aviv's cybersecurity pipeline: gaming, military, and startups 08:57 Why AI/agents are fundamentally hard to secure (new OS + 'human spirit') 13:44 Trust, autonomy, and permissions: managing the blast radius 18:17 Real-world exposure: Claude Code and the developer identity attack surface 20:16 A workable approach: treat agents as untrusted processes with identity + least privilege 22:33 Zero Trust for Agents: Access ≠ Permission to Act 23:27 Token's "Intent-Based Permission Management" Explained 25:29 Building the Identity Map: Tracing What Agents Touch 26:52 The Secret Sauce: Using AI to Secure AI in Real Time 28:10 Real-World Case: 1,500 Agents and Wildly Over-Provisioned Access 30:57 CUA 'Computer-Use' Agents: Exciting, Personal… and Terrifying 34:44 Secure-by-Default & Sandboxing: Fixing 'Always Allow' Dark Patterns 35:36 What Security Teams Should Do Now: Inventory, Boundaries, Governance 37:59 What's Next: Agent Teams and Multi-Day Autonomous Work 40:10 Tony Stark Vision: Agents That Improve the Human Experience 41:02 RSA Innovation Sandbox: Token's Big Bet on Intent + Access 43:01 Wrap-Up, Audience Q&A, and Sponsor Message