Microsoft unveils Guardian Agents to monitor AI agents — TechCrunch

Microsoft has unveiled Guardian Agents, a new feature for monitoring AI agents used by companies. It is part of Project Perception, a cybersecurity project introduced in June, TechCrunch reports.

Risks from AI agents

Project Perception was initially created on the assumption that AI tools for hacking posed an external threat. The project had three types of agents: a “red team,” which acted as malicious actors and carried out attacks; a “blue team,” which detected the actions of the former and assessed the vulnerabilities found; and a “green team,” which was meant to eliminate them.

Microsoft is now also emphasizing the risks posed by agents that enterprises themselves connect to their systems and data. Microsoft CEO Satya Nadella said in his blog that companies should not rely solely on model providers’ assurances about the security of agentic products. In his view, such systems should be monitored and controlled by separate tools that are not under the control of AI model developers.

More current news is available on the UA.News Telegram channel Telegram.

New security features

Microsoft’s security business chief, Hanyatte Gallo, said that frontier models, regardless of their origin, cannot be considered fully reliable because of their unpredictability. According to her, models are trained to perform tasks rather than follow organizational norms or demonstrate proper judgment.

Microsoft has also added an agent detection feature to Microsoft Defender. It can find up to 60 types of agents that employees bring into the work environment, including Claude Code, Codex, and OpenClaw. Other new capabilities include limiting agents’ access to data and systems, dynamically adjusting permissions depending on the task, and requiring human approval if an agent is likely to make risky decisions.

The company has also become one of the first participants in Nvidia’s Open Agent Safety Platform, which has been backed by more than 100 technology companies. The platform isolates agents in protected environments and can track and stop their behavior using hardware.