Trust Is the Deployment Layer
Agents fail when their access boundaries are unclear as well as when their answers are wrong.
OpenAI previewed Private Safety Processing this week: a way to monitor longer AI interactions for abuse while keeping eligible enterprise data under stricter retention rules.
That control still needs operating rules around it.
Frontier labs now have to sell trust as part of the product. Agents read files, call tools, cross systems, remember context, and touch the parts of the business people used to keep out of demos.
That changes the adoption question.
What can it see?
Who approved that access? Where do the logs live? Who reviews the output before it becomes a decision? What data is never allowed into the workflow at all?
If those answers live in procurement paperwork, they are too far away from the work.
Privacy belongs in deployment architecture as well as the vendor review. The companies that define those boundaries can move. The ones that cannot will keep pausing every time an agent gets close to real data.
The gap is knowing exactly what your agents are allowed to know.