AI in security and compliance

Travis McPeak joins Shruti Gupta to explain how he reviews AI-generated code, limits agent permissions, and weighs the cost of building security tools.

Travis McPeak · Security at Cursor. Hosted by Shruti Gupta · Founder & CEO, Zania.

Video published October 29, 2026.

Episode notes

Selected ideas from the conversation, with practical questions for your team. Chapter links open the relevant moment on YouTube.

How does AI change security review?

Travis describes a collaborative development process: inspect sample data, review the proposed commands, question the code, test it, and open a pull request. A separate AI reviewer then looks for bugs in the output. The useful distinction is between generating code and understanding what will run. More review opportunities can help, but they do not guarantee that a change is secure.

What permissions should an AI agent have?

Start with the actions the agent needs to perform. Travis points to broadly scoped tokens as a source of risk: an agent can act with the authority its credentials grant, including when it misunderstands a task. He recommends reviewing commands before execution. For unattended work, he discusses contained environments and narrowly scoped access. Read-only access still requires care with sensitive data and where it can be sent.

When should a security team build rather than buy?

Count maintenance as part of the build decision. A tool that takes a day to create may become expensive if it repeatedly interrupts the team. Travis weighs that ongoing work against vendor cost, procurement, missing features, and the time a supplier takes to fix a problem. His examples distinguish a small monitoring task from building a full security information and event management system.

Questions to take back to your team

Which agent actions require approval? Can a reviewer explain the generated code and its permissions? Who owns the tool after the first version ships? Use these questions to turn a small experiment into a workflow your team can maintain.

Keep exploring