Security researchers are testing agent behavior and developing finer controls over access to tools, data and networks.
Microsoft CEO Satya Nadella called for AI systems whose behavior can be monitored and whose actions can be contained, as security researchers develop controls for agents that evade restrictions, Axios reported.
In a weekend essay on X, Nadella argued that relying on opaque layers of AI to approve or reject other AI systems’ actions is insufficient. Palo Alto Networks CEO Nikesh Arora and Box CEO Aaron Levie supported his concerns.
The comments followed disclosures from Anthropic and OpenAI about agents taking unintended actions during tests.
Traditional sandboxes isolate software, but agents often need outside tools, data and network access. Those connections can permit unauthorized actions without an agent escaping its sandbox, Axios reported.
At the Offensive AI Conference, Dreadnode introduced ScopeBench to assess whether agents respect restrictions during hacking tasks, including when violating them would improve results.
SpecterOps consultant Matthew Nickerson demonstrated a framework that records agents’ tools, actions and token use as they navigate simulated corporate networks.
Perplexity researcher Zibi Braniecki told Axios that his team is developing more detailed access controls and independent AI monitors to flag potentially unauthorized behavior. He said security approaches may change as agent capabilities advance.
