Google Research published a report on October 5 examining unresolved security questions for autonomous AI agents. The researchers warn that human approval can become less effective during long tasks and when agents delegate work to other agents.
Repeated requests can lead users to approve actions without inspecting each one. The access appropriate for a task can also change as work progresses. An existing permission does not establish whether a later action still fits the user's request.
The authors propose an additional control layer. It would check planned actions against the current task and context, then allow them, block them or ask for clarification. Reliably generating and enforcing those rules remains a research challenge.
The report grew out of a November 2025 workshop. It also examines how safeguards can survive cooperation between multiple agents. These proposals set out a research agenda rather than demonstrating that the problems have been solved.