Microsoft said on the 20th it has extended Project Perception, a security platform built to counter external AI attacks, to monitor and control AI agents working inside companies. The idea is to treat AI agents as untrustworthy insiders and check them continuously.
According to the Kyunghyang Shinmun, Microsoft said on the 20th it widened the scope of Project Perception so it can detect AI agents running inside a company and bring them under management. The core is assessing the risk of actions an agent is about to take and supervising it continuously. Until now AI security has focused on stopping outside attackers who use AI models.
광고 문의 · 300×250The platform includes a ‘Guardian Agent’ team that watches every agent action around the clock and responds in real time when risk rises. After a risky action is caught, it provides traceable data so the cause can be identified. Figures such as pricing or the number of customer firms were not confirmed in the coverage.
Why watch ‘internal AI’
Hayete Gallot, a Microsoft security executive, wrote on the company blog that agents are widely adopted, can access data and systems and make decisions at key steps, so they are effectively insiders that need monitoring and security. The aim is to end the habit of trusting agents without verification simply because the company deployed them.
CEO Satya Nadella said on the 11th, local time, that providing information and controlling information should be separated. He argued that control mechanisms must sit outside the AI model and that a model should not act as its own security guard. The Kyunghyang Shinmun added that Microsoft has criticized Anthropic’s publication of its ‘Claude constitution.’
Open questions
For companies, the more agents they run, the more permission management and audit logs become new security tasks. Who is accountable when an agent malfunctions or is abused, and who verifies the monitoring tools themselves, remains unsettled. No concrete standards or regulatory moves on this were confirmed in the coverage.
This should not be confused with surveillance of human employees. The announcement concerns the behavior of software agents, not staff. Still, because agents can access employees’ work data, privacy concerns may be worth readers’ attention, though that is not something confirmed in this coverage.