← sources

An AI agent shouldn't have more permissions than the employee it's assisting.

source: reddit (reddit:AI_Agents)published 2026-10-23 12:16ingested 2026-10-24 02:35

retained text

Imagine deploying an AI agent to help your sales team.

It needs to read customer information, update CRM records, and create follow-up tasks.

So someone gives it broad API access because that's easier than configuring permissions for each action.

Now the agent can potentially modify records it shouldn't touch, access unrelated customer data, or execute actions outside its intended role.

This is a familiar security problem, but AI agents introduce a new way for it to surface.

I'd start with least-privilege permissions, scoped credentials, validation of tool inputs, and explicit approval for sensitive actions.

I'd also treat retrieved documents and user messages as untrusted input, not instructions that can override the agent's permissions.

An agent making a bad decision is one problem.

An agent having the authority to turn that decision into a serious incident is another.

How are people handling authorization for agents that interact with multiple business systems?

document fetched 2026-10-24 10:22 · canonical https://reddit.com/r/AI_Agents/comments/1xdce8a/an_ai_agent_shouldnt_have_more_permissions_than