← sources

How much access should an AI agent have?

source: reddit (reddit:AI_Agents)published 2026-10-27 08:41ingested 2026-10-28 02:35

retained text

I keep running into this question when thinking about agents connected to real tools.

A workflow might start with read-only database access. Then it needs an API, email, or an MCP tool to finish the job. Each addition seems reasonable, until the agent can read sensitive information, change records, or contact people.

That is why I do not think a system prompt can be the security boundary. Telling an agent “do not access this” does not help much if its credentials still give it access.

For people running agents beyond experiments, what safeguards are actually working for you?

document fetched 2026-10-28 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xgn4rp/how_much_access_should_an_ai_agent_have