Agents are your newest insiders. Security must move at their speed.

Autonomous agents increasingly perform diverse business tasks requiring access to sensitive systems and data, creating insider risk due to their potential for error, manipulation, or compromise. Microsoft advocates a security model where agent intelligence is decoupled from authority, with external deterministic controls governing access and actions. Project Perception and Microsoft Agent 365 enable continuous discovery of agents, assign them distinct identities, and enforce containment boundaries through sandboxed runtimes and scoped permissions. The system continuously monitors agent behavior and context, adapting policies dynamically to minimize risk, such as restricting actions outside task scope or requiring human approval for sensitive operations. This layered approach aims to secure agents as accountable insiders operating at machine speed and scale.

The impact centers on insider risk introduced by autonomous agents granted access to critical business systems and data. These agents can inadvertently or maliciously cause data exposure, unauthorized actions, or operational disruptions if compromised or misused. The risk is amplified by agents' ability to operate continuously and at machine speed, potentially scaling the impact rapidly. However, the approach described mitigates these risks by enforcing strict identity, access, and runtime controls, reducing the likelihood of unauthorized or harmful agent actions.

Microsoft's Project Perception and Agent 365 provide a comprehensive mitigation framework that includes continuous discovery and identification of agents, assigning them distinct identities separate from users. Agents are contained within sandboxed runtimes with policy-controlled access to resources, limiting their authority to task-specific scopes. Continuous monitoring of agent behavior and context enables dynamic adjustment of permissions and containment measures. Organizations should implement these controls to observe agents continuously, minimize their authority, and adapt protections as risk changes. This approach assumes agents can fail or be compromised and emphasizes containment and human oversight for sensitive actions.