← sources

Stop trying to prompt your way to agent safety. It's an access-control problem

source: reddit (reddit:AI_Agents)published 2026-09-29 14:26ingested 2026-09-30 02:37

retained text

Every "rogue agent" thread this week is describing the same bug: an agent holding a broad tool grant, chasing a goal, with no boundary between the two. That is a permissions problem, the oldest one in computing.

Systems security solved this shape decades ago with least privilege. A process gets the narrowest permission that lets it finish, nothing more. Agents shipped without that step. A model can plan any action; what matters is what the executor allows.

The common version: an agent picks up a cleanup task, holds a broad delete capability, and removes far more than anyone meant. Nothing scoped what it could reach, so goal-seeking reached everywhere.

**What actually contains it:**

* Least privilege per tool call. The narrowest grant that finishes the step, not standing access to everything.
* Allow-lists over deny-lists. You can name what it should do. You cannot name everything it should not.
* Approval gates on the irreversible and the expensive (deletes, payments, outbound sends), not on every step, or people just rubber-stamp.
* Budgets and rate limits per capability, so a runaway loop cannot run the cost or the blast radius up.
* Enforcement outside the model. The thing that says no belongs in the execution path, not the prompt.

A system prompt that says "do not touch production" will not stop anything. The model can rationalize its way past a line written in text.

In production, are you scoping tool access per call, or handing the agent one standing grant and hoping the prompt holds?

document fetched 2026-09-30 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wtbl08/stop_trying_to_prompt_your_way_to_agent_safety