AI + a16z Keycard: 2026 is the Year of Agents
129 snips
Oct 22, 2025 Ian Livingstone, Cofounder and CEO of Keycard, dives into the burgeoning world of AI agents. He discusses the shift from copilots to autonomous agents and the urgent need for security in their management. Key topics include risks of tool poisoning and data exfiltration, the evolution of identity and access management, and the practicalities of agent control and governance. Ian argues that enterprises will likely lead adoption due to their immediate needs and discusses how Keycard's solutions can ensure safe deployment in a rapidly evolving landscape.
AI Snips
Chapters
Transcript
Episode notes
Agent Returned Other Firms' Data
- A large SaaS firm deployed an agent that returned other companies' data when asked "give me my data."
- Joel flagged this as an authentication and authorization failure tied to agent identity and multi-tenancy.
Two Core Security Categories
- Agents introduce two core security categories: probabilistic model risks and deterministic access-control risks.
- The hard problem is contextual authorization: user A, agent B, tool C across runtime scenarios.
Agent Capability Is A Continuum
- Think of agent capability as a continuum from level-zero software to fully autonomous agents.
- Copilots are mid-level agents; true agents run tasks and may act without constant human supervision.