Skip to content
MaungaAI news, explained

Safety and security

Microsoft says don't trust AI agents, and sells the watchdog

Microsoft warns that companies' AI agents can act like a planted employee. Its solution is other agents to watch them, and Microsoft is the one selling them.

2 min read2 sources, one officialWritten with AI · machine-translated

Share: WhatsApp · Email

AI-generated illustration for Maunga. Not a photograph.

The essentials

  • On October 19, Microsoft unveiled Guardian Agents, agents that oversee every action taken by a company's other AI agents (Microsoft Source, TechCrunch).
  • Claimed by one partyHayete Gallot, Microsoft's head of security, says no leading model is reliable: not because it is malicious, but because it is unpredictable (Microsoft Source, TechCrunch).
  • Claimed by one partyAccording to TechCrunch, Microsoft Defender can now detect up to 60 types of agents that employees bring to work, such as Claude Code or Codex (TechCrunch).
  • Claimed by one partyAccording to TechCrunch, the system matches each agent's permissions to its task and asks a human for approval if a decision looks risky (TechCrunch, Microsoft Source).

Why it matters

One of the big AI companies is telling its customers not to trust the models it sells and funds itself. This suggests that agent security is shaping up as a separate business, a bit like an audit: the watcher kept apart from the maker. TechCrunch finds the argument sound, even if it suits Microsoft (TechCrunch).

What does it mean for me?

If you work at a company that uses AI agents, or you install one on your own, the IT department could start detecting them and limiting what they can touch.

The details

The new features expand Project Perception, the security system Microsoft unveiled this summer. It uses three teams of agents: one attacks like a hacker, another finds the flaws and a third fixes them (Microsoft Source, TechCrunch). Microsoft dates it to July, and TechCrunch to June (Microsoft Source, TechCrunch). It was built to stop attacks coming from outside. Now it also targets the agents already working inside the company (Microsoft Source).

According to TechCrunch, Microsoft CEO Satya Nadella already called on companies in October, on his blog, to monitor agents with systems that the AI companies themselves don't control (TechCrunch). He summed it up as the need to “separate the supply of intelligence from authority over it” (TechCrunch). The outlet points out that Microsoft has invested in OpenAI and Anthropic, and that it wants to be the one selling that oversight (TechCrunch).

What we don't know

  • How much these features cost and when they will reach all customers, including in Europe.
  • Whether anyone independent has checked that the watchdog agents really stop the ones that go off track.
  • Whether the watchdogs run on the same models they watch, and so share the same blind spots.

Sources

This story is based on 3 sources. In the text, each fact names in brackets who reports it.

Was this story useful? Yes Not really

Spanish original