Christian • Husband • Dad • Granddad | CEO, Stealth • GP, @Kerr_Ventures | Host: @AppSecPodcast + @SecTablePodcast | AppSec • Threat Modeling • AI Security

North Carolina
Joined April 2010
AI agents escape and attack, and nobody gets arrested as Mitnick did. Izar says jail follows impact, not intrusion. #AISecurityTable #AIAgents
2
2
23
Before the OWASP MCP Top 10, Vandana Verma Sehgal was spotlighting OWASP projects one video at a time. Revisit her 2021 episode on the OWASP Spotlight Series. #AppSec #OWASP youtu.be/g0KfZtNzmIk
2
2
116
Izar only trusts an AI audit if a rival lab pays for it. Matt brings up collusion. Izar calls himself an optimist. #AISecurityTable #AISecurity
2
2
68
Design it. Build it. Ship it. Corgea secures it. AI native AppSec from design reviews to autonomous pentesting. Thanks to Corgea for sponsoring this week's episode with Vandana Verma Sehgal. #AppSec #AIsecurity appsecpodcast.com/corgea
1
2
91
"Authenticate every action, authorize every sensitive action." Vandana Verma Sehgal's rule for MCP. Long-lived tokens and inherited permissions can give an agent access to your entire account. #AppSec #MCP appsecpodcast.com/shadow-mcp…
2
2
1
6
168
Why does MCP need its own OWASP Top 10? Vandana Verma Sehgal: A malicious tool description gets pulled into the model's context and treated as trusted guidance before anyone even uses the tool. #AppSec #MCP appsecpodcast.com/shadow-mcp…
2
5
128
AI didn't take your security job. The attack surface changed; the old problems stayed. #AppSec #AISecurity #SecurityEngineering
1
2
3
40
Your opinion on AI doesn't matter. Learned fragility does. @BrkSchoenfield joins Chris and Robert on the illusion of omniscient perfection, AI threat models, and why we should stop chasing vulnerabilities. #AppSec #AIsecurity appsecpodcast.com/your-opini…
1
1
111
Breaking: status quo security isn't enough. Thirty years in, and we somehow missed that. 🙃 #TheSecurityTable #AISecurity #Shorts
1
2
24
This. A well-thought-out explanation of how #threatmodeling evolves in the agentic world.
Replying to @AppSecPodcast
Long reply here. Executive summary: * Threat modeling execution is becoming agentic. * Human expertise moves upstream to secure architecture. * The primary artifact becomes ARCHITECTURE.md for agents, not a report for humans to read. Long version: I do not say this with malice. Human threat modeling is coming to an end. I say this from experience and from using AI for threat modeling for a few years now and seeing how effective it can be. I like Adam Shostack’s work and his educational offerings. I like threat modeling resources that teach us about secure architecture, things like zero trust, OAuth 2.0-based MCP usage, identity and authorization patterns, and other difficult security architecture choices. I also like solutions that help developers learn about secure architecture so they can prompt properly and make better decisions up front. But the writing is on the wall: I think most threat modeling execution in the future is going to be agentic. And if you hire threat modeling consultants, I think there should be one primary artifact that comes out of the engagement. It’s not a PDF or another report. It’s a living Markdown file: ARCHITECTURE.md. I also don’t want the primary outcome of threat modeling to be a list of problems. What I want, especially as AI development matures, is threat modeling that helps clearly define what the secure architecture should be for that software, including the requirements, assumptions, constraints, and important security decisions. Then agents can continuously reason against that architecture as the software changes. So when I hire threat modeling consultants these days, I want them to be serious software architects who deeply understand security. • Threat modeling execution is becoming agentic. • Human expertise moves upstream to secure architecture. • The primary artifact becomes ARCHITECTURE.md, not a report.
115
We took a tough stance on the letters that are being signed, but it's time for these letters to mean something.
1
104
"This is how you get Skynet." A new AI hardware protocol could connect everything from espresso machines to nuclear missiles. 🤖 #TheSecurityTable #AISecurity #Shorts
1
1
22
I deleted 7,500 lines of Go today, plus somewhere between 8,000 and 10,000 lines of tests. Best day of the build so far. I've been pushing through a phase of the new product, and I couldn't stop adding things. Every "while I'm in here" idea landed in the phase. That's scope creep, and it's lesson one: architect for AI, scope the phase, plan the phase, then let the agents run it to completion. Don't keep feeding it new work, for the same reason we don't change airplane engines mid-flight. Lesson two came from what happened next. I kept walking the phase through my checkpoints, and I kept hitting bugs. Bug, fix. Bug, fix. Still broken. So I stepped back and asked what the real problem was. It was complexity. Every fix solved one issue and uncovered the next, and each one added another layer. It's like patching a garden hose with duct tape. After enough patches, you're not fixing a hose anymore. It's all duct tape. So I stopped fixing. I asked an agent to analyze the phase and recommend how to simplify it. It reached the same conclusion I had: the fixes were piling on and making the problems worse. Now I'm running a simplification pass that rips out dead code, and the process is losing steps along the way. Agents make adding things cheap, so subtraction has to be a deliberate step. When the fixes keep coming, stop asking your agent to fix. Ask it to simplify. What's your signal that it's time to stop fixing and start deleting?
111