Speaking Topic: Agentic AI Risk & Governance
The liability surface nobody is auditing yet, because the risk is not in the code, it is in what the agent is allowed to do.
I talk about agentic AI risk as a 2026 diligence blind spot, because most technical review processes still ask whether the code is correct, when the real question is what autonomy the system was granted and whether that autonomy was ever stress-tested against adversarial or simply careless input.
In forensic audits, I have found rogue agent liability hidden inside AI stacks worth 50 million dollars or more in enterprise exposure, not because the model was wrong, but because nobody had defined what the agent was not allowed to do, and the agent found the gap faster than the engineering team did.
This is not a hypothetical future risk. It is the same class of failure as an ungoverned service account, except the account can now write its own permissions requests in plain English and sound convincing while doing it.
Key Takeaways
- 01Code review is not an AI agent liability audit; they test different things entirely.
- 02Undefined agent boundaries are a measurable, auditable risk, not a philosophical one.
- 03Agentic AI risk belongs in the same diligence category as access control and permissions review.
Frequently Asked
How is an AI agent liability audit different from a code review?
A code review checks correctness. An AI agent liability audit checks what autonomy the system was granted and whether its operating boundaries were ever adversarially tested.
From the Q&A
“Doesn't a capable model reduce the need for this kind of audit?”
No, because the risk is almost never model capability. It is undefined authority boundaries, which a more capable model can exploit faster, not more safely.
“What's the first thing you check in an agentic AI deployment?”
What the agent is actually authorized to write to or modify, versus what its designers assumed it would be limited to.