I’m back: practical AppSec guidance on AI agents, human accountability, secure delivery, and programs that work.

Hi there,

After a two-year break, Reasonable AppSec is back.

I’m Chris Romeo. I’ve spent my career building application security programs, companies, communities, and conversations about making software more secure. Today, I’m building again as a repeat cybersecurity founder—and paying close attention to how AI is changing both software development and application security.

That’s why I’m bringing this newsletter back.

Reasonable AppSec will focus on the decisions behind effective security: secure design, threat modeling, developer workflows, AI-assisted development, and the difference between buying another tool and building a security program that actually works.

My goal is simple: to curate the AppSec and AI stories worth your attention, explain why they matter, and give you a practical perspective you can use.

Thanks for sticking around. I’d genuinely like to know what AppSec problem is taking up most of your attention right now. Hit reply and tell me.

P.S. I use AI to help organize this newsletter. But whenever you see “Chris’s Take,” those words are mine—bespoke, old-school, and typed from the heart of fifteen years of experience in AppSec

Some things stay the same, and some things change. I’m a late bloomer in the field of AI and security. I took a sabbatical after the sale of Devici to play on my farm and plant a garden. What a time to take a break! I came back to this whole new world where coding agents have taken over development. AI changes the pace of development, not the responsibility. As AppSec folks, we are still responsible for the security and privacy of the code that goes into our products. It’s just different now.

With this brave new world, there is more tension between faster software delivery and continued human accountability. Agents increase the velocity of code deploys and pull requests. As humans, we’ll fight to stay in the loop, but we have to find a way to do this without squashing the velocity. This is the same problem security has always faced. There has always been tension between security and dev because security wants to slow things down while dev wants to build cool stuff. The AI part of the equation just cranks up the pace at which security must excel to keep up. If we try to slow this thing down, we’re going to lose, and five years from now, AppSec will be something that used to be part of dev but got squashed because it couldn’t keep up.

Agents are operating mostly autonomously now. We can act like they aren’t, but the reality is that they don’t stop and ask for confirmation on everything they do. And they shouldn’t, because that would make them nowhere near as useful. This is why it’s important for us to know what the agent was allowed to do in some way—and what it actually did. We need a control that sits between the agent and the world so that, at a minimum, we can audit their actions. Hopefully we can block its actions as well. Visibility into permissions and actions matter if we want to call what we’re doing “secure development.”

Treat agent-generated code as an untrusted contribution to your repo at this stage. Evidence must be generated and reviewed for everything the agent is doing. People use the term “intern” or “junior developer” to describe how they see an agent today. We wouldn’t let an intern push to prod on their first day. We would take them down the path, but a serious code review would be conducted on their PR. AI output still requires validation, review, and clear ownership.

We want to move fast. That is not in doubt, but speed without verification accelerates security debt. We’ve been racking up security debt with a human squad for decades now. We have a chance in this brave new world to use the tools as they continue to evolve to clear out our security backlogs. The catch is that we can’t just turn them loose and hope for the best. Hope is never an executable plan.

Productive automation is our path forward. Tools that work create a partnership between development, security, and AI. DevSecAI? Dare I try to coin a new term in an industry buried in four-letter acronyms? DevSecAI is a faster path to reducing this backlog, but its success depends on the principle that no delivery can be unreviewed.

The Future of Open-Source Threat Modeling
AI is showing up in the threat-modeling conversation whether we like it or not. We get practical about where it helps, where it creates false confidence, and why human judgment still owns the risk.

Make No Mistakes: Inside the First “Agentic Ransomware”
The “agentic ransomware” label is getting attention. We pull it apart, look at the evidence, and ask the question that matters: was this actually a new class of attack or familiar automation with a new wrapper?

What did I miss this week? Hit reply and tell me.

— Chris