Skip to content
Menu

Agent permissions / PRAXENOR PERSPECTIVES

More agents should not mean more permission

If an AI agent delegates a $100 task, its helpers should share that $100. Creating helpers should not increase what you approved.

Share one allowance across the team

Consider a hypothetical agent authorized to spend up to $100 on one task. If it creates four helpers and each gets a new $100 allowance, the team could spend more than the person approved.

Our research explores a rule in which the whole team shares one approved allowance. A helper uses part of that allowance instead of getting a new one. In the design we are testing, creating helpers also uses part of the team’s remaining allowance.

The dollar example makes the idea concrete. It is not a claim that every permission can be reduced to money or a single risk score.

Keep money, data and publishing limits separate

A spending cap, permission to read a folder, and authority to publish a document are different constraints. Unspent money should never unlock an otherwise forbidden action.

OWASP’s excessive agency guidance recommends limiting tool functions and permissions, independently enforcing authorization, and requiring human approval for high impact actions. Our proposed shared allowance would add a limit across the team. It would not replace checks on each action.

Prevent helpers from spending the same money

Suppose two helpers each request $60 while $100 remains. Both could see sufficient funds if the system only reads the balance before execution. The system must reserve or subtract funds before allowing the requests, so they cannot both spend the same money.

Decide what happens when a request fails, is cancelled or runs again. Does a failed request put the reserved money back? Can the system tell a repeated request from a new purchase? Test these rules as well as the displayed balance.

People keep control of permissions

In our proposed design, a monitoring tool flags concerning patterns. It cannot give itself permission to disconnect agents or remove their access; a person must approve those responses. An existing permission check can still refuse a prohibited action immediately.

Test this with a fictional task involving one main agent and two helpers. Try requests at the same time, repeated requests, a forbidden action and a new helper. Check that the team stays within the approved limit and that creating helpers never creates new permission.

Praxenor’s work on shared permissions is still research. Tests of a prototype are a starting point; they do not prove protection in real use.

Continue with how to review activity across an AI team. For implementation questions, read technical details and check current product status.

← All security insights