Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’

Microsoft has introduced security features to monitor enterprise AI agents, warning that even powerful “frontier” models cannot be trusted to act predictably. The shift reflects concern that agents with access to sensitive data and business tools could become insider risks, so companies need independent oversight of their actions.

The main addition, Guardian Agents, is part of Microsoft’s Project Perception, launched in June with agents designed to simulate attacks, detect vulnerabilities and help fix them. Microsoft security chief Hayete Gallot said models are trained to complete tasks rather than follow an organisation’s norms or exercise its judgement. The article notes Microsoft’s commercial interest in selling agent security products, while arguing that independent monitoring can still make sense.