Trust, But Verify: Understanding What AI Agents Actually Do

Abstract:

AI agents are quickly moving from assistants that generate answers to autonomous systems that take action. They now execute commands, access files, call APIs, interact with SaaS applications, and make decisions on behalf of users.

As security practitioners, how do we trust an agent when we can't easily see what it is actually doing?

Knowing an agent's identity, authenticating the user who initiated it, and defining what the agent is authorized to access are important but not everything. An agent can have legitimate credentials and legitimate permissions while still taking an unexpected, risky, or malicious action.

The next frontier of AI security may therefore be less about simply asking "Who is this agent?" and more about asking "What is this agent trying to do, and is that behavior consistent with what we intended?"

In this webinar, we'll explore the emerging discipline of agent behavior security. We'll discuss how to connect identity, intent, context, and runtime actions to establish trust in autonomous systems.

We'll explore how to make agents observable, explainable, and governable so organizations can move from experimenting with AI agents to confidently deploying them at scale.

Topics we'll explore:

• Identity vs. behavior: Why knowing who an agent is doesn't necessarily tell you whether its next action is trustworthy.
• Intent vs. execution: How an agent's original goal can diverge from the actions it ultimately takes.
• The runtime as a security boundary: Why commands, file access, API calls, MCP tools, and other agent actions may be the most meaningful points for security decisions.
• Context matters: How user, device, repository, application, tool, destination, and action context can change whether the same agent behavior is safe or risky.
• From visibility to enforcement: How organizations can start by understanding normal agent behavior and progressively introduce guardrails without stopping adoption.
• Building trust at scale: What an enterprise-ready model for agent governance could look like as agents become more autonomous and gain access to increasingly sensitive systems.