AI agents on your credentials, not on an admin account

The most common security mistake in an AI agent is not in the model. It is in the permissions.

The typical recipe is well known: create a service account with admin rights so the agent "can do everything". It works in the demo. Everyone claps. And it is an incident waiting for a date.

Why the admin account is a trap

An agent with admin permissions can touch your whole Vault, whether it needs to or not. One misread prompt escalates into a data incident. And that service account becomes a permanent backdoor nobody audits, with access to more than any individual user should have.

The problem is not hypothetical. It is structural: you gave the least predictable component of your stack the highest permissions in your system.

The opposite decision

When we built VaultPilot together with Avant Leap, the first native AI agent for Autodesk Vault Professional 2026, I made the opposite decision: every chat runs with the user's own credentials. No admin backdoors.

In practice it means this: if you cannot release a drawing, neither can the agent. If you cannot edit that property, neither can the agent. If your role cannot see a project, the agent cannot see it either. The agent inherits your limits instead of bypassing them.

Your PDM security remains exactly what your administrator configured, with the same rules for humans and agents. There is no second attack surface to manage.

Permissions are not the whole story

On top of the permission model go the other layers: dry-run with confirmation before every write, validation against the real Vault schema, and a full audit trail for every operation, with who, what and when.

An agent that respects your PDM security is not less powerful. It is the only one you can let into production. When you evaluate any "AI for engineering", start with this question: which permissions does it run with?