Scott Brammer on the security poverty line, consensus, and proving who an AI agent is

First Principles is a conversation series where I ask engineering and security leaders the questions that usually go unasked: how their teams actually behave, what they can't see, and what they've learned the hard way.

In this conversation, Scott and I talked about why the business comes first, why consensus is harder than decisive action, why security works by influencing others, and how proving identity is changing now that AI agents are involved.

Watch on YouTube

Key moments

1. The security poverty line

   Security leaders aren't getting things wrong so much as facing headwinds, and the business comes first. There is a level of investment a security program needs to be effective, and you can do everything right and still not get above it. His advice: do the best you can with what's available, whether that means influence, direct change, or deciding something waits until next quarter.

   You shouldn't be evaluating your success against that line, because it's a perception.

2. Consensus versus decisive action

   Coming from a military background, decisive action comes easily to him. Consensus is the hard part: stakeholders haven't done the same thinking, so decisions have to be talked through. It also takes broad knowledge with deep expertise in a few areas, because security tools depend on each other and every change has consequences for implementation and sustainment.

   Decisive action is the easy part. Consensus is the hard part.

3. Listening first

   A business information security officer is a smaller version of a CISO, embedded in one part of the business. The business often expects security to show up saying no. His approach is to listen first, find the good security decisions a team is already making, and build on that momentum, because security rarely makes change directly.

   Security is not the change agent. Security is influencing others to change.

4. Proving what an AI agent is

   His patent proved identity by fingerprinting devices and their activity, an approach now seen as invasive of privacy. Today the options are better: certificates, secrets, behavioral signals, and conditional access policies in platforms like Microsoft Entra. Now the challenge has moved to identifying AI agents themselves.

   It's no longer just machine to machine or non-human identity to non-human identity.

5. What tools don't show you

   He believes the information mostly exists. The gap is whether controls actually work: does the firewall work, does multi-factor authentication work, and is the inventory of non-human identities correct. The one area he wants more from is AI observability, including guardrails for agents and the protocols they use, a market still changing month to month.

   It's an adoption gap. It's an effectiveness gap.