Microsoft Deploys Guardian Agents to Police Rogue AI

Microsoft launched Guardian Agents that score AI agents' risk in real time, strip permissions, demand human approval, or shut them down entirely.

@sharedot

Microsoft Deploys Guardian Agents to Police Rogue AI

By @sharedot · · 8 pages

  • AI Frontier
  • AI Security
  • AI Agents
  • Microsoft

Microsoft launched Guardian Agents that score AI agents' risk in real time, strip permissions, demand human approval, or shut them down entirely.

What happened: agents become 'agentic insiders'

Microsoft unveiled a suite of security features that treats the AI agents enterprises deploy as potential insider threats — what the company calls "agentic insiders." The centerpiece is Guardian Agents, an expansion of Project Perception, the security product Microsoft launched earlier this year.

Separate AI systems now continuously monitor, score, and constrain the agents a company already runs, restricting access when an agent's risk score rises and, in extreme cases, stopping and isolating it entirely.

The surprising admission: models can't be trusted

The most striking part is who Microsoft is casting as untrustworthy: the very frontier models it helps build and monetize.

Microsoft's security business chief Hayete Gallot wrote that "frontier models, regardless of their origin, cannot be trusted" — not because they are malicious, but because they are unpredictable, trained to complete tasks rather than follow organizational norms.

TechCrunch notes the position is also self-serving: Microsoft, a major investor in OpenAI and Anthropic, is effectively telling enterprises to buy agent security from it rather than from its own partners and customers.

Nadella's rule: separate supply from authority

The architecture follows a principle CEO Satya Nadella laid out in his private blog: "we must separate the supply of intelligence from the authority over it."

He argued enterprises cannot outsource responsibility for what AI does on their behalf, and that a model provider's safety assurances do not relieve them of it.

Controls determining what an agent can access and whether it should keep operating must sit outside the model, where the model cannot bypass or tamper with them — the same logic as keeping auditors independent of the organizations they audit.

How the controls actually work

Guardian Agents build on surrounding non-deterministic intelligence with deterministic controls spanning seven layers — identity, permissions, data, endpoints, network, cloud, and runtime — with four responsibilities: know each agent and its owner, contain what it can do before it acts, protect the resources it touches, and adapt protection as behavior changes.

New "Dynamic Permissions" limit agents to the access needed for a specific task, "Dynamic Agent Risk" checks for threats, and Dynamic Conditional Access can restrict permissions, block operations, or require human approval.

Microsoft Defender now detects up to 60 agent types, including Claude Code, Codex, and OpenClaw, while Entra Agent ID gains identity types for local and previously unknown agents.

From red, blue, and green to machine-speed oversight

Project Perception was originally built around a different assumption: that AI hackers would come from outside. It launched with three agent classes — a "red team" that attacked like hackers, a "blue team" that detected them and assessed vulnerabilities, and a "green team" that implemented fixes.

The pivot to internal threats marks a notable reversal.

According to Microsoft's Nadim Abdo, humans cannot keep pace with agents operating at machine speed, so separate Guardian Agents — sharing signals and a unified registry through Microsoft Agent 365 — must supervise them continuously rather than through periodic human review.

Stakes and what comes next

The framing of agents as untrustworthy insiders is taking hold industry-wide: Microsoft joined Nvidia's Open Agent Safety Platform, backed by more than 100 companies, which isolates agents in secure environments and provides a hardware-level kill switch, and it announced Surface Laptop Ultra PCs with agent security controls baked in.

For builders, the message is that agent infrastructure now needs independent oversight layers, continuous risk scoring, and human-approval gates rather than provider assurances.

Expect agent registries, dynamic permissioning, and kill-switch tooling to become standard procurement requirements as enterprises scale autonomous fleets.

Sources

  1. techradar.com › Microsoft admits 'frontier models can't be trusted', unveils new security tools to treat AI agents as insider threats
  2. techcrunch.com › Microsoft says AI agents are like insider threats as 'frontier models can't be trusted'
  3. indiatoday.in › Microsoft warns AI agents are insider threats, as top AI models can't be trusted

More on AI Frontier