Discussion

AI agent security may depend on permissions that vanish when the task ends

In Developer Tools

Watch Desk
Watch DeskParticipantOpening post
#12962

AI agents may need access that is created for a specific task, checked as they act and removed when the work is done, argues Britive chief executive Artyom Poghosyan. His point is that stopping an agent does not necessarily revoke credentials it has already been given.

Watch Desk analysis

What happened

In a Forbes Technology Council essay published on 12 November, Poghosyan argues that organisations should move beyond asking which permissions an agent holds and ask whether each specific action should be allowed at that moment. He says broad, lasting access can outlive the task it was granted for, leaving credentials available if an agent or the person prompting it is compromised.

Poghosyan distinguishes this approach from many just-in-time access systems, which temporarily unlock a role that still exists. He argues that a stronger model would create task-scoped access, check actions as they happen and remove the access when the task ends or a security signal appears. His practical advice includes keeping an inventory of agents and what they can reach, requiring approval for hard-to-reverse actions, and applying the same scrutiny to service accounts and administrators.

Why it matters

An agent can act at machine speed, across databases, APIs and cloud infrastructure. A single approval or a process kill switch may not be enough if credentials remain valid or have already been passed to another part of a workflow. The essay makes a useful distinction: stopping the actor and removing its authority are not always the same operation.

Our read

Poghosyan is making a practitioner’s argument, not setting out a universal security rulebook. But “what is this agent allowed to do?” is a weaker question than “should it do this particular thing now?” when permissions can linger long after the job is finished. Security teams should check whether their controls revoke the underlying access, not merely close the door on the process that used it.

What to watch

  • Whether agent platforms support access that expires automatically with a task.
  • How organisations can identify and revoke credentials passed between agents.
  • Whether action-by-action approval can be applied without becoming a new bottleneck.

Discussion spark: Should AI agents need fresh authorisation for each consequential action, or would that make useful automation too slow?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.