Microsoft’s security leaders published the doctrine on 19 October

On 19 October 2026 Microsoft published two posts that read more like a warning to customers than a product launch. Hayete Gallot, executive vice president of Microsoft Security, wrote “Securing the agentic insider” on the Official Microsoft Blog. Nadim Abdo, corporate vice president of Identity and Network Access, published “Agents are your newest insiders. Security must move at their speed.” TechCrunch’s Julie Bort covered both the same day, at 3:04 PM Pacific, which is after midnight in Kyiv, so readers in Europe and Ukraine met the story on 20 October.

Gallot’s claim is blunt: “Unfortunately, frontier models, regardless of their origin, cannot be trusted.” She adds that agents “are unpredictable. They will adapt as the context changes, may not behave as intended and will collaborate and collude with other agents to reach their objectives, without needing to be compromised.”

The collusion sentence is the strongest thing in either post, and it needs a caveat. It appears only in Gallot’s piece; Abdo does not repeat it, and neither does Microsoft’s 16 October post with Nvidia on open-weight model risk. No incident data or customer case is attached to it, so treat it as an unsupported vendor assertion about future behaviour.

The doctrine: separate identity, no standing access, enforcement outside the model

Abdo’s post is where the argument becomes architecture.

An agent should not wear its sponsor’s identity. Abdo writes that an agent should not “automatically inherit all the standing access of its user” and should “begin with no authority and receive only what it needs for the approved task, for only as long as the task requires.” Microsoft is introducing Dynamic Permissions to handle that grant, plus Dynamic Agent Risk and Dynamic Conditional Access for agents.

The second principle generalises beyond Microsoft: “the model cannot be its own security boundary.” The controls that determine what an agent can reach, what it can do and whether it keeps running “must sit outside the model, where the model cannot bypass or tamper with them.” That is correct, and it is the free part of the doctrine. Prompt instructions and model-level guardrails are guidance to a probabilistic system. They are a poor enforcement boundary, and they should not be the only thing standing between an agent with standing access and production systems.

Satya Nadella framed the same idea on his scratchpad on 10 October: “we need to separate the supply of intelligence from the authority over it.” Abdo restates it for agents in the 19 October post.

Discovery is activity based, and that limits the inventory

Microsoft’s discovery feature now covers 35 additional types of local agents, “bringing the total to 60 supported local agents, including Claude Code, OpenAI Codex, and GitHub Copilot.” The full list is the Defender for Endpoint AI agent support matrix, which also names OpenClaw. Microsoft’s post and TechCrunch each name a different subset, so treat those names as examples; the matrix holds the full list.

Microsoft is explicit that inventory is not control: “Discovery is activity based. Defender lists an agent after it observes agent activity on an onboarded device. Installing an agent doesn’t by itself guarantee that the agent appears in the local AI agent inventory.” Abdo draws the line himself: “Discovery tells you that an agent exists. Identity gives security and governance systems a durable anchor for controlling it.”

The published limits are narrower than the announcement’s tone. Network inspection “doesn’t support agents that use certificate pinning or HTTP/3”, per the runtime-protection overview, and the support matrix notes that runtime protection support “doesn’t indicate that protection is enabled on a device.” An agent that never runs on an onboarded device stays invisible, and one that pins certificates or speaks HTTP/3 will not be seen on the wire.

The licence boundary the 19 October posts leave out

Neither post states a licence requirement. The boundary sits in the Entra Agent ID documentation, last updated 24 June 2026: Entra Agent ID “is available for all Microsoft Entra customers,” but “extending Microsoft Entra security features to agents requires Microsoft Agent 365. Agent 365 is included with Microsoft 365 E7 and is available as an add-on to Microsoft E5/A5/Business Premium (or Microsoft Defender Suite + Microsoft Purview Suite).”

Runtime protection is preview. The Defender runtime-protection page carries a preview label, as does the Windows runtime-protection section of the support matrix, and local agent discovery has preview labels on the macOS and Windows Subsystem for Linux tables. The main Windows discovery table is not labelled preview.

Guardian Agents, the supervisory agents Gallot describes, exist as an announcement only. They sit inside Project Perception, which Microsoft announced on 27 July 2026 and placed in public preview on 3 August 2026. There is no general availability date and no named outside evaluator.

So “get started today” comes with three qualifications: the doctrine is free to read, the Entra Agent ID platform is open to Entra customers, and the security features grafted onto agents need Agent 365. Which Defender plan covers agent discovery and runtime protection is not stated on the pages Microsoft published, so put that question to a licensing contact.

The company selling the monitoring has stakes on both sides

TechCrunch frames the launch as both surprising and self-serving. Surprising, because Microsoft “is famously close to OpenAI and Anthropic, having invested in both. OpenAI is also one of its biggest cloud customers, and Microsoft has its own AI lab developing models.” Self-serving, because Microsoft, in TechCrunch’s summary, “wants enterprises to buy agent security products from it, the cloud provider, rather than from Anthropic and OpenAI.” TechCrunch also offers the analogy that agent monitoring should be separate from the suppliers of agents, just as auditors should be independent of the organisations they audit.

Keep those two ideas apart. The principle holds without the conflict: whoever supplies the model should not be the only party checking it. Microsoft’s own posts never mention its OpenAI or Anthropic positions.

Terrell Cox, CVP and Deputy CISO for Customer Security, says Microsoft is “already operating in a world of hundreds of thousands of agents” as Customer Zero, with no published figure, methodology or audit to support the claim. Microsoft calls Guardian Agents “an established industry term,” and no non-Microsoft source we could reach uses the phrase that way. No independent evaluation of Guardian Agents exists in the material available to us.

What to adopt now, and what to check before buying

  1. Give every agent its own identity, so its permissions and its activity logs belong to the agent itself and no longer roll up under the human who started it. Entra Agent ID does this and works with non-Microsoft agents through a sidecar SDK or workload identity federation.
  2. Start at zero authority. Issue task-scoped, time-bound grants and let them expire. Standing access is what turns a misbehaving agent into an incident.
  3. Enforce outside the model. Run agents in a sandbox with policy-controlled access, such as the Microsoft Execution Container (MXC), where security controls sit outside the agent’s reach. Check whether your agents pin certificates or use HTTP/3.
  4. Keep an inventory and assume it is incomplete. Activity-based discovery finds agents that run on onboarded devices, and it will miss an agent that never triggers telemetry.
  5. Require human approval for consequential steps: production writes, credential changes, payments, data exports. An agent that can approve its own high-impact actions is the insider problem with more speed.
  6. Before buying, check preview status and the Agent 365 requirement. Do not promise a control to auditors until both are confirmed. Microsoft Ignite runs 17 to 20 November 2026 and is likely to move some of these labels.