Microsoft on Monday introduced new security features aimed at the risks of AI agents. The main addition is Guardian Agents, which monitors the AI agents a company uses. It forms part of Project Perception, a Microsoft security product launched in June.
Microsoft has also added an agent detection feature to Defender. It can find up to 60 types of agent that employees bring into work, including Claude Code, Codex and OpenClaw. Other new features limit what agents can access and adjust their permissions to suit the task. They also require human approval if an agent appears to be making risky decisions.
Hayete Gallot, Microsoft's security business chief, said frontier models 'cannot be trusted'. She said this is because they are unpredictable, not because they are malicious or easily hacked. Models are trained to complete tasks rather than follow organisational norms, and they often have wide access to data and tools.
Chief executive Satya Nadella has made a similar case. He argues that businesses should not rely on the assurances of model providers and should keep control of agents separate from the firms that supply them.
What it means for your business
Find out which AI agents and coding tools your staff already use, as many may have been adopted without approval. If you run Defender, check with your reseller or in your admin portal whether these features are included in your licence. Whatever tools you use, limit what each agent can reach and require a person to approve any action that changes data or spends money.
This brief was written automatically from the source above and summarises it. It is not legal, tax or financial advice.