tag / security

Browser-resident AI agents inherit the human's authenticated session by design. The threat model that follows is not prompt injection. It's session theft with consent, against an identity stack that assumes the browser is the user. A short note on what DPoP, CAEP, and the still-unfinished sender-constraint story actually fix.

A POC pitted an agent-based compliance scanner against an active network scanner on the same Windows domain controller. They disagreed about whether it was vulnerable to BEAST. The disagreement is the whole story.

SPDX and CycloneDX round-trip on paper. In practice the same SBOM produces different vulnerability findings depending on which scanner reads it, because the two formats disagree about identity, scope, and license in ways that change what a vulnerability database can match against.

Windows TLS configuration is split across a registry tree, a per-process .NET override, and a runtime cache. Each layer reports the truth about itself and lies about the others. A short field note on how teams convince themselves a domain controller is hardened when it isn't.

Microsoft has been moving LDAP channel binding from 'available' to 'required' for nearly seven years. The enforcement timeline keeps slipping because the third-party tail is large and quiet. A short note on what the registry key actually does, how to find the appliances that will break, and what to do before the DC stops accepting them.

Passwordless authentication is finally landing across enterprise IAM programs. The same orgs are simultaneously onboarding AI agents that hold long-lived credentials and run inside the human's session. The two projects are colliding, and most identity programs aren't ready.

Token exchange is the cleanest pattern for delegating identity to an AI agent, and every major IdP implements it differently. Field notes from wiring it into Entra, Okta, and Auth0 — what the spec leaves to interpretation, what breaks first, and what to standardize before the second agent ships.

AI agents are the next vulnerability vector and we are not tackling it fast enough. A look at why the gap between AI adoption and AI governance is widening, what it looks like inside real orgs, and where identity programs should be aiming.

After more than a decade in enterprise security across the public and private sectors, I'm convinced the actual perimeter is human and machine identity. Most security programs are still spending the majority of their budget on the wall around the castle.

I built a small security lab to teach the failure modes I see most often in client work. The first real test was running it against my own production site.

The orgs most exposed to AI-era security risk are the ones moving fastest to ship with AI. They are also the orgs least likely to have a CISO. I built Diallo Security Advisors for that gap.

I came up through racking servers, hypervisor migrations, and identity provisioning. The lesson that work taught me, and that I keep applying to AI in 2026, is that you cannot make good security calls from one set of binoculars.