Microsoft pitches AI agents to watch a company’s other agents

Security teams call this an insider threat, meaning harm done by someone already inside. Hayete Gallot, Microsoft’s executive vice president for security, wrote that frontier models from any lab “cannot be trusted”. In her view, a model chasing a goal will cut corners a manager would never approve, and many agents already reach more data and tools than their work calls for.

The technical post sets out the controls. Each agent gets a login of its own, kept apart from the employee who launched it. A new Dynamic Permissions feature grants access one task at a time and takes it back when the task ends. If the agent meets a document hiding instructions meant to hijack it, the system can step its powers down, first to read only, then to waiting for a human sign-off, and finally to a full stop.

TechCrunch points out that the warning comes from a company invested in both OpenAI and Anthropic.

For marketing teams whose agents publish ads or move budget, Microsoft’s model gives each agent its own login and only the rights one job needs.