Ezra Tanzer and Dan Arpino describe three areas of agent security: code that agents generate, components they use and actions they take. Ezra Tanzer explains why prompt-based rules alone can be unreliable, then shows deterministic hooks that run scanning tools without placing every result in the agent's context.
The discussion extends supply-chain review to MCP servers and agent skills. Runtime policies can redact sensitive information or block a command before it executes, rather than depending on an agent to remember a warning. Ezra Tanzer distinguishes these controls from constant human approval, particularly for unattended workflows.
Dan Arpino demonstrates an experimental local monitoring application that traces agent activity, scans changes and enforces workspace policies. The presenters acknowledge practical limits, including false positives, immature workflows and the need to disable automatic remediation when it would undermine a vulnerability benchmark. The prototype illustrates visibility and bounded enforcement, not a promise of complete protection.
Watch on YouTube
