Microsoft argues that AI agents running inside companies should be treated as high-risk insiders because they are often overprivileged. The company describes how Project Perception and 'Guardian Agents' will discover agents, give them an identity and policies, and dynamically score the risk of each action. Depending on that risk, an agent can be restricted, made to wait for human approval or isolated entirely.

What it means: The announcement shows a shift in how the big vendors think about AI security. Until recently the debate was mostly about attackers using AI from outside. Now Microsoft openly says that agents a company deploys itself can cause damage, not out of malice but because they chase results without grasping the organisation's rules. Its proposed answer resembles what we already do for people: a separate identity, minimum permissions for a limited time, and monitoring of actions. The difference is that checks must happen in real time, before an action runs. For Greek businesses trying agents in accounting, customer service or software development, the takeaway is practical: before an agent gets access to payment systems, ERP or customer data, there should be a record of what each agent does and a way to stop it immediately. GDPR and NIS2 requirements make that traceability even more important. Microsoft also calls for more transparency from the models themselves, so that controls can judge not only what an agent is doing but why it is doing it. That would make containment far smarter.

Read the original on Microsoft

Related service: Custom software

Related stories