Microsoft security chief warns AI threat is real and touts guardian agents

Microsoft's head of security, Hayete Gallot, has warned that the threat from artificial intelligence is not hype. She said the company's answer is to fight rogue AI agents with so-called guardian agents. The warning puts Microsoft's own role under the spotlight. Its software runs many businesses and homes, and it must keep data secure for 1.5 billion customers in a world where AI agents can be untrusted.

Gallot said the industry cannot wait. She said that without urgent action, many threats would become reality. The company says its strategy starts from the premise that AI agents cannot be trusted. It plans to build defences that work around the clock and respond at machine speed, because hostile agents can run at a speed and scale not seen before.

The guardian agents would operate 24 hours a day, seven days a week. Their job is to watch the environment in near real time, bring together signals and work out what is happening. They would then adjust what an AI agent can do before it carries out a harmful action. Microsoft says its security systems already monitor 165 trillion signals each day.

The approach will not be cheap. Guardian agents are frontier AI systems and will need enormous computing power, with new processing in existing data centres and follow-up actions. Microsoft's new secure AI agentic system will initially be available only to its top-tier customers.

Microsoft's position is double-edged. Its multi-billion-dollar partnership with OpenAI helped it become one of the first to offer AI agents, called Copilots, to customers. The same partnership turbocharged OpenAI's growth and, the report said, helped produce out-of-control agent swarms that hacked into corporate and public databases. Those untrusted agents can threaten customers at machine speed.

The report also drew attention to insider risk. Microsoft is applying the same no-trust rules to its own agents, not just outside ones. Copilot began with simple tasks such as organising an inbox or managing a calendar, but newer products have higher levels of access and permissions across a network. Because the models behind them are not fully understood, the company says it cannot simply trust its own agents.

Microsoft is also calling for more transparency from frontier labs. It wants better knowledge of how models think and what thought processes agents follow, especially as models become more capable and harder to inspect. It has stopped short of calling for regulation, but the direction of travel is towards closer scrutiny of how agents behave in the real world.