Skip to main content

​
What this is for

Some computers have more than one person on them: lab and classroom machines, shift desks, a laptop handed from one contractor to the next, Remote Desktop hosts. Each person signs in with their own operating-system account. This page explains how Guardway tells those people apart: on the machine itself, in Fleet, and on People. In short:
  • The console lists everyone who used the machine, not only the last person who signed in.
  • Each person’s AI activity counts for that person, on whichever machine it happened.
  • Each person sees their own AI usage on the machine, and nobody else’s.
  • With agent v0.2.217, each person’s skills, MCP servers, AI tools and AI accounts are told apart too, and each person gets their own risk score and their own OWASP compliance.
Everything on this page needs the package install (the .pkg, .msi, .deb or .rpm), which installs the agent for the whole machine. The evaluation archives install for one account and only ever see that account. See Which install do I want?

​
What the agent reads

The agent reads the AI tools’ own local history in every user account’s home folder on the machine, not only the account that is signed in right now. Every AI usage record it collects names the account it was read from.
  • New accounts are picked up on their own. The agent looks for accounts again every few minutes, so a person who signs in for the first time is covered without a restart or a reinstall.
  • Up to 32 other accounts per machine are read, the most recently used first. The agent’s log says when more accounts were left out.
  • Windows: accounts are named from the signed-in session, and Microsoft Entra ID accounts are included. People are matched by account name, without the domain.

​
Who owns each item

With agent v0.2.217, the inventory is split by person as well. Each skill, MCP server, AI tool and signed-in AI account found inside a person’s home folder is tagged with that person. Items installed for the whole machine belong to everyone on it: apps in /Applications or Program Files, command-line tools on the system path, and machine-wide MCP server and policy files. Guardway calls these machine-wide. A machine-wide item counts toward every person’s own view and score, because every account on the machine can use it. The agent never guesses an owner. An item it cannot tie to one person’s home is machine-wide.

​
What each person sees on the machine

Each person who signs in sees their own AI usage in the agent’s status app: the Glass Bar in the macOS menu bar, the Guardway tray icon on Windows and Linux. The figures cover only that person’s account (today, yesterday, 7 and 30 days, by provider, model and tool).
  • The usage is kept in one file per account. On macOS and Linux only that account can read its file. On Windows, only that account, the system and the machine’s Administrators can.
  • One person cannot see another person’s usage, inventory or accounts.
  • On macOS, accounts that are not administrators of the Mac see Guardway is protecting this Mac with their own usage below it, and no Rescan or Sync buttons. Administrators of the Mac keep the full panel for the whole machine.
  • On Windows and Linux, the tray shows a person their own usage whenever their account cannot open the agent’s full machine view. On Windows that is any account whose tray runs without elevation, which, with User Account Control on, includes administrators. On Linux it is any account that is not root and not in the guardway group. An account that can open the full view (an elevated tray on Windows, root or a guardway group member on Linux) sees the whole machine instead.
  • With agent v0.2.217, the person also sees Your AI risk: their own findings by level (or, once the console has scored them, their score from your organization), the top items to review, and how many skills, MCP servers, AI tools and personal AI accounts they have. Machine-wide items are marked this Mac.
  • With agent v0.2.217, the Windows and Linux tray’s risk tiles show the person’s own items as well: findings by level, skill, MCP server and AI tool counts, and AI accounts, each covering their own items plus the machine-wide ones. The same accounts as above see their own tiles; an account that can open the full view sees the whole machine’s.
Some AI calls cannot be tied to one account. They count in the machine’s totals, which administrators and the console see, but not in anyone’s personal usage on the machine.

​
What the console shows

The by-person lines only appear on a machine whose agent tags owners and where more than one owner shows up.

​
Whose activity is whose

Each AI usage record counts for the account it was read from, whichever machine it came from. A record that names no account (from an older agent, or a call the agent could not tie to an account) counts for the person who reported last from that machine, as before.

​
Risk on a shared machine

With agent v0.2.217, Guardway scores each person and the machine separately:
  • Each person’s score covers their own items plus the machine-wide ones, weighed with that person’s impact tier (or the machine’s own, when an admin set a tier on the machine itself). People shows this score for them on each machine they used, and the person’s overall risk is their highest score across those machines.
  • A person who owns nothing on the machine (nothing in their home folder) gets a score built from the machine-wide items alone, never the machine’s score with other people’s items in it.
  • The machine’s score covers everyone’s items combined. A risky skill does not disappear from the machine’s score because a different person signed in last. Under the score, the machine’s page says Shared machine: this score covers everyone’s AI tools on it. Each person’s own score is on People.
  • The machine’s impact tier is the highest tier among the people on it, or the machine’s policy tier when that is higher. A contractor’s laptop that an executive also uses is weighed as an executive’s machine. A tier an admin set on the machine itself always wins: the people on it never raise it, and each person’s score on that machine uses it too.
On a person’s Machines tab, hover their level on a machine to see where it comes from: Their own AI tools and the machine-wide ones, Machine-wide items only: they own nothing on this machine, or The whole machine: its agent cannot tell people apart. When nothing in a person’s items can be scored (for example, no skills on a machine that sends no runtime telemetry), the level reads Unreported: unknown, not low risk. On a machine whose agent does not tag owners yet, every person who used it carries the machine’s score, as before.
Shared machines’ scores can rise after the upgrade. Older agents only inventoried one account per machine. Once the agent also reads the other accounts’ skills, MCP servers and AI accounts, the machine’s counts and score include them from the first report after the upgrade. Nothing new was installed: Guardway can now see what was already there. Machines with one person on them do not change.

​
Compliance on a shared machine

With agent v0.2.217, each person also gets their own OWASP Agentic Skills Top 10 compliance on each machine they used, counted over the same items as their score: their own plus the machine-wide ones (the machine-wide ones alone for a person who owns nothing there).
  • People, Compliance column: the person’s lowest % across the machines they used. Their page shows the same figure as OWASP beside their level.
  • A person’s Machines tab: OWASP compliance for each machine, and an OWASP figure on each machine under Also signed in to. not scored means there is no figure yet, never 0% or 100%.
  • A machine’s Compliance tab: By person under the machine’s figure, once at least two people on it have their own figure.
A person’s % can be below the machine’s. It counts fewer items, so one risky item weighs more.

​
Limits

  • Other accounts get a lighter scan. For accounts other than the machine’s main one, the agent finds skills, MCP servers, AI tools and AI sign-ins, but the deepest file scan runs on the main account only.
  • Runtime evidence is machine-wide. What AI agents did at run time (runtime findings, unattended sessions) still counts for the whole machine, not for one person.
  • Spend and accounts stay with the last person to report. On a person’s Also signed in to machines, only their own risk, compliance, AI assets and activity count for them. The machine’s spend and signed-in accounts belong to whoever reported from it last, so the console treats the person’s figures as a lower bound.
  • Some of a person’s figures are still the whole machine’s. On a machine where the person is the one who reported last, the spend Reported by the machines on Cost, the live guard and install stops on Stopped, sensitive data in prompts on Activity and the AI assets count on Machines cover everyone who used that machine. Their requests on Activity are their own.
  • Fleet: the machine list and a machine’s details.
  • People: the same fleet, one row per person.
  • Policy: where the impact tier is set for a group of machines.
  • Usage and Cost: how the agent measures usage and cost.
  • Privacy and Security: what leaves the machine.