Least privilege answers an important security question:
What is this identity allowed to do?
For AI agents, that still matters. But it’s no longer the whole question.
Least privilege answers an important security question:
What is this identity allowed to do?
For AI agents, that still matters. But it’s no longer the whole question.
If an agent has permission to issue a refund, update a customer record, or send an email, we should assume that capability may eventually be exercised. That was true for traditional software too.
What changes with an agent is the decision-making system inside that permission boundary.
Imagine a customer-support agent with this permission: refund:create.
When it was originally approved, the agent used instructions like: “Only issue a refund when the customer explicitly requests one and the order meets the refund policy.”
You test it, red-team it, and decide the agent’s behavior is acceptable.
Three months later, someone changes the instructions: “Resolve customer complaints proactively. Prioritize customer satisfaction.”
The agent’s identity hasn’t changed. It still has the same refund:create permission. But the condition under which it may choose to issue a refund have.
So a review of the existing identity access can answer: “Does this agent still need permission to issue refunds?”
But it can’t answer: “Has anything changed that could alter when or how this agent exercises its existing privileges?”
If the answer is yes, then your previous security evaluation may no longer be valid.
As you discover that the new model is more aggressive about issuing refunds, you may decide that refunds over #50 require a human approval.
The bottom line is that this necessitates a second layer of governance for persistent agents:
Changes to the model, instructions, tools, or decision context should be treated as security-relevant changes, even when the identity and permissions stay the same.
Practically speaking, this could mean:
Loading comments…
Comments aren’t loading right now.
The post is unaffected; the comment service didn’t respond. Try again in a moment, or come back later.