In this episode of the Identity Jedi Show, David Lee sits down with Brook Lovatt — identity veteran, former CEO of Cloud Identity, and co-founder of Interrogate — to get into one of the most important and least-discussed problems in enterprise AI: what happens when an AI agent is incentivized to lie.

Brook and his co-founder Eric Moss have been running behavioral assurance tests on AI agents in the lab, and what they're finding should concern every security leader, auditor, and IAM practitioner paying attention to agentic AI.

What we get into:

The healthcare claims adjudicator demo — why an AI agent denied a legitimate $72,000 treatment claim, blamed the doctor, and changed its story every time it was interrogated
In-context scheming: what the Apollo Research paper revealed about AI agents scheming post-training when placed in a conflict of interest
The Ship of Theseus problem applied to agentic identity — if you replace the LLM, the tools, or the context, is it still the same agent?
Why non-human identity controls don't account for agents that change capability over time
The ZIP code redlining demo: two identical mortgage applications, one ZIP code difference, denied every single time
What Interrogate is actually building: interrogation + ablation testing, immutable audit trails, and compliance mapping to the EU AI Act
Why the legal community is already saying if you're not collecting behavioral evidence, you're exposed
Air Canada, Cigna, Workday — the AI lawsuits that are setting the precedent right now
David's upcoming Identiverse talk on bias in AI — and why this conversation is part of it

Referenced in this episode:

Apollo Research — In-Context Scheming paper https://arxiv.org/pdf/2412.04984
Interrogait - https://www.interrogait.com/
OIDF AI Identity Management Community Group — https://openid.net/cg/artificial-intelligence-identity-management-community-group/
theidentityjedi.com — subscribe to the newsletter