Microsoft Copilot
M365 Copilot
Microsoft's AI assistant built into Microsoft 365 apps such as Word, Outlook and Teams. It works with the files and data the signed-in user can already access.
In our briefings
The claim each of the newest briefings on Microsoft Copilot would have you leave with.
A security buyer no longer takes the category claim as evidence. A hospital system's security head gets from the whole market to a purchase by must-have lists, analyst shortlists and peer references down to fewer than ten demos, then three or four proofs of value run inside his own environment, and the vendor that behaves like a partner and ships a dashboard the sponsor can read wins. The vendor side of this corpus already said the demo closes the deal; the buyer says the demo is where his evaluation starts.
Vendors all claim to cover AI agents and non-human identities, so buyers test them in their own environmentBI-FIND-2026-0465
For hosted agents such as ChatGPT, Claude.ai and Microsoft Copilot, where the customer controls no network path, code or host, Errico's AARM spec has one full route: the vendor runs a hook synchronously before every side-effecting tool call and hands the action to a policy service the customer controls. "If the vendor does not offer hooks, AARM cannot be deployed", he writes; the fallback is enforcement at the customer's own APIs, which covers only the tools the customer exposes. Even with a hook, bypass resistance "depends entirely on vendor implementation". AARM defines no wire protocol for the hook. Errico calls this a collective action problem: vendors have little reason to build hooks until customers demand them, and customers cannot demand them without a standard to point to, which is the role he proposes for AARM.
An open spec for stopping agent actions before they run needs the vendor's help to cover ChatGPT or CopilotBI-FIND-2026-1223
Feng, McDonald and Zhang separate an agent's agency—its capacity to act through tools—from its autonomy, which they define as "the extent to which an AI agent is designed to operate without user involvement," and treat the two as "two distinct levers for agent governance." Holding the model, tools and desktop environment fixed, they map five user roles in Table 1: L1 Operator (user plans and invokes; Microsoft Copilot, ChatGPT Canvas), L2 Collaborator (shared planning and takeover; OpenAI Operator), L3 Consultant (agent leads and elicits preferences; Gemini Deep Research, Replit Agent, GitHub Copilot Agent), L4 Approver (user steps in only on failures, missing credentials or consequential actions; SWE-agent, Manus, Devin), and L5 Observer (logs and an "emergency off-switch" only; Voyager, The AI Scientist). At L4, they warn that prompting users for passwords and API keys raises security risks from storing credentials, alongside the risk of disengaged users rubber-stamping approvals.
Washington researchers split agent autonomy from tool access and propose third-party autonomy certificatesBI-FIND-2025-0441
Cited in 4 briefings
Every briefing that discusses this term, newest first.
- Vendors all claim to cover AI agents and non-human identities, so buyers test them in their own environment
Podcast · · The Non-Human & AI Identity Podcast · NHI & AI Identity Summit : The Buyer's Dilemma · Sep 7, 2026
- An open spec for stopping agent actions before they run needs the vendor's help to cover ChatGPT or Copilot
Paper · · arXiv · Herman Errico · Feb 10, 2026
- Washington researchers split agent autonomy from tool access and propose third-party autonomy certificates
Paper · · arXiv · K. J. Kevin Feng et al. · Jun 14, 2025
- The 2023 paper that named indirect prompt injection showed an attacker can use whatever access the assistant has
Paper · · arXiv (cs.CR) · Kai Greshake et al. · Feb 23, 2023
Related terms
What appears alongside Microsoft Copilot in the same briefings.
AgentforceGitHub CopilotMicrosoftActive DirectoryMicrosoft Entra Agent IDCursorSalesforceChatGPT