Microsoft launches tools that treat AI agents like insider threats
Microsoft has released new security tools that watch the AI agents companies use at work. A Microsoft security chief wrote that frontier AI models "cannot be trusted" because they are unpredictable. The tools can limit an agent's access when it strays from its job.
Microsoft on Monday released new security features meant to protect businesses from the AI agents they use, treating those agents like possible insider threats, TechCrunch, Help Net Security and TechRadar reported.
The main addition is Guardian Agents, which monitor the other AI agents a company runs. It is part of Project Perception, a security system Microsoft introduced in June, according to TechCrunch. That system used three kinds of agents: a "red team" that attacks like hackers, a "blue team" that finds threats and a "green team" that fixes problems.
"Unfortunately, frontier models, regardless of their origin, cannot be trusted," Hayete Gallot, the head of Microsoft's security business, wrote, according to TechCrunch. She said that is not because the models are malicious, but because they are unpredictable and are often given access to sensitive data and tools.
The new controls can change an agent's permissions as its behavior and risk change, Help Net Security reported. Microsoft is also expanding its tools to spot 60 types of agents that run on people's computers, including Claude Code, OpenAI Codex and GitHub Copilot. Each agent gets its own identity, separate from the person using it.
"We should not ask a model to be its own security authority," Nadim Abdo, a Microsoft vice president for identity and network access, wrote in a company blog, TechRadar reported.
Microsoft CEO Satya Nadella has made the same argument. "We need to separate the supply of intelligence from the authority over it," he wrote earlier this month, according to TechCrunch.
TechCrunch called the move both surprising and self-serving. Microsoft has invested in OpenAI and Anthropic and builds its own AI models, but it also wants businesses to buy its agent security products.
What we know: Microsoft has launched Guardian Agents and new controls that limit AI agents' access based on risk.
What we don't: What the tools will cost and how widely businesses will adopt them.
Where outlets agree TechCrunch, Help Net Security and TechRadar all report that Microsoft's Guardian Agents extend Project Perception to watch and contain AI agents as possible insider risks.
Where they differ TechCrunch credits the "cannot be trusted" warning to security chief Hayete Gallot, while TechRadar says the announcement blog was written by Nadim Abdo.