I tested 130 sign-in cases. MFA was often not the control that decided it.
130 conditional access, MFA and AI agent cases, a 72-page book, and the scripts now run in your browser.
Hello,
Last issue was about machine identities. This one is about the other half of identity: how people and AI agents sign in, and where conditional access, MFA and single sign-on stop working.
Same method as before. I wrote every guess first, hashed the file and committed it before any code existed. Everything is made-up data, and I wrote both the guesses and the models, so a match proves little. The misses and the assumptions are the useful part.
The numbers
The lab has 130 cases. I ran 124 on small models. 118 matched my locked guess and 6 did not, all in the run where two or three controls worked together. Six cases need a real tenant, a hardware key, a real agent product or a person on the phone, so they are written up for anyone to run.
MFA was often not what decided it
A stolen integration token never triggers a sign-in, so MFA plays no part and an address allow-list decides the case. A password reset that does not end refresh tokens leaves a stolen session alive. A recovery path weaker than sign-in lets an attacker enrol their own authenticator.
AI agents
Ten of the cases are about agents. The controls that decided them were plain ones: a named human owner, an identity that expires with the task, and a check at the resource against the person's own rights. Rotating an unused agent's credential changed nothing, because rotation hands the agent a new working credential. A scan that flags an ownerless agent does not remove its access.
One assumption at a time
When I changed one assumption and ran again, results flipped. The one I would show a security team first: continuous evaluation caught a cookie stolen after MFA until the attacker looked like they were on the victim's network. After that, only binding the session to the device held.
Run it yourself
The scripts now run in your browser. Change an assumption and see which cases change. Change one word in a guess file and the script refuses to run, because the hash no longer matches.
roshantrivedi.co.in/ca-mfa-lab/virtual-lab/run/
Read more
The book, Who signs in? (72 pages, free): roshantrivedi.co.in/ca-mfa-lab/ca-mfa-lab-book.pdf
The write-up: roshantrivedi.co.in/writing/130-conditional-access-mfa-and-ai-agent-cases/
The framework, seven links from enrolment to revocation: roshantrivedi.co.in/ca-mfa-lab/framework/
One question for you
If you had to pick one of the seven links to check in your own setup this month, which one would it be, and why?
Reply and tell me. If you test a case in a real product and it behaves differently from my model, there is an exception form on the lab, and your result goes on the page next to mine.
Roshan