Oct 7, 2026 · 1 hr 20 min · 17 segments
**Introduction** - Vito's career beginnings at JD Edwards - Founding of Nextworld sparked by a reunion conversation - The evolving role of platforms in enterprise software **Interview** - Vito's…
Michael BernzweigHostAditya GanjamGuestArvind ParthasarathiGuest
Alexis MoyseGuestZiad BashirGuestUh, today we're gonna be talking about AI security and governance, and ultimately that means trust.
Uh, my perspective is that you need to trust your system of record and let that serve as the basis for governance and trust, and not necessarily the model or the LLM I believe a lot of what enterprises are being told is that AI security starts with the model layer, that we should work to make the model more trustworthy.
My position is that trust doesn't necessarily belong with the model, but that it belongs in the system the model runs within, which is your enterprise platform or your system of record.
Let's, let's discuss, uh, an example or a possible incident, a what, what if, if you will.
So we have an agent, and it was asked to clean up some customer dataset, um, and it did just what it was asked.
It ran against production because the account it executed under could reach the production environment.
Every ev- individual step was the kind of thing you would have signed off on yourself.
So where did things go wrong? Three days later, the data owner asks why a specific set of accounts was collapsed, changed, and no one can quite answer because the only record of the run is some chats, chat chan- transcript that is already scrolled out of view.
Look closely, and there are two failures here, and neither one is about the quality of the model.
The first is that the agent could reach data it should not have been able to reach, and the second is that no one could reconstruct why it acted the way it did.
So the agent had reach it shouldn't have had, and there was no evidence of intent and decision-making, and that's where the problem lies.
Uh, today we're gonna be talking about AI security and governance, and ultimately that means trust.
Uh, my perspective is that you need to trust your system of record and let that serve as the basis for governance and trust, and not necessarily the model or the LLM I believe a lot of what enterprises are being told is that AI security starts with the model layer, that we should work to make the model more trustworthy.
My position is that trust doesn't necessarily belong with the model, but that it belongs in the system the model runs within, which is your enterprise platform or your system of record.
Let's, let's discuss, uh, an example or a possible incident, a what, what if, if you will.
So we have an agent, and it was asked to clean up some customer dataset, um, and it did just what it was asked.
It ran against production because the account it executed under could reach the production environment.
Every ev- individual step was the kind of thing you would have signed off on yourself.
So where did things go wrong? Three days later, the data owner asks why a specific set of accounts was collapsed, changed, and no one can quite answer because the only record of the run is some chats, chat chan- transcript that is already scrolled out of view.
Look closely, and there are two failures here, and neither one is about the quality of the model.
The first is that the agent could reach data it should not have been able to reach, and the second is that no one could reconstruct why it acted the way it did.
So the agent had reach it shouldn't have had, and there was no evidence of intent and decision-making, and that's where the problem lies.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.