#agent permissions

A monitoring agent on every machine is a lot of trust to ask for. This page is the exact scope of that trust: what the agent reads, what it cannot do at all, and what has to be approved before it happens.

##Read-only by default

A freshly enrolled agent cannot change anything. It reports state and waits. That is not a hardening setting you have to find; it is the install.

desk-04 · agent capabilityread-only

$ infotechbang-agent capability mode read-only reports inventory, patch level, health, software list can execute nothing pending actions 0 outbound agent.infotechbang.com:443 (tls 1.3) inbound none · no listening port an action appears here only while an approved one is in flight

The agent opens no listening port and accepts no inbound connection. It polls out over TLS. There is no path from the internet to a device through it, because there is nothing on the device to connect to.

##What the agent reads

The complete list, not a summary of it.

Category	Collected
Identity	Hostname, serial number, model, manufacturer, platform and build number
Hardware	CPU, memory, disk size, disk health counters, battery cycle count, network interface names and MAC addresses
Software	Installed package names and versions as the platform's own inventory reports them
Patch state	Installed and missing updates with their vendor severity
Licence	Product keys and activation state only where the vendor exposes them locally
Session	The signed-in account name, to attribute the device to a person. Not what that account is doing
Reachability	Check-in timestamps and uptime

File contents are never read. The agent sees that a package is installed; it does not open the files that package created.

##What it never does

These are not disabled defaults. The code to do them does not exist in the agent, which is a materially different claim and the one worth making.

No screen recording or screenshots. No screen capture API is linked.
No keystroke logging. No input hook is installed.
No clipboard access.
No personal file access. Documents, photos, downloads and desktop contents are not enumerated or read.
No browser history, bookmarks or saved credentials.
No camera or microphone.
No location or geolocation, including for a device flagged as lost.
No network traffic inspection. No packet capture, no TLS interception.
No remote desktop and no remote shell. There is no mechanism to get a prompt on a user's machine.
No productivity or activity scoring. Idle time is not reported.

Tell your staff this, in writing

The list above is deliberately quotable. Rolling an agent out without saying what it does creates a suspicion that is far more expensive than the rollout. A short note on the day, naming what is read and what is not, settles it.

##Approved actions

An action is a named, bounded operation on a named device. It requires an admin approval that states the scope before it exists, and it expires if not approved.

action envelopewhat approval actually authorises

action        service.restart
target        desk-04                # this device only
parameter     spooler                 # this service only
requested by  ticket #2231 suggestion
approved by   ravi.k · 18 Sep 2026 09:21 IST
expires       15 minutes if unapproved
state before  spooler: running, queue 14 jobs
state after   spooler: running, queue 0 jobs
result        ok · 1.2s

###The action catalogue

There is a fixed list. An admin cannot write an arbitrary command, and neither can we - which is the control that matters, because it means the worst case is bounded by this list rather than by imagination.

Restart a named service
Clear a print queue
Install an approved patch from a vendor channel
Uninstall a named package
Restart the device, within its maintenance window
Revoke identity sessions for a named account
Disable or re-enable an account in the linked identity provider
Run a named, read-only diagnostic and return its output

No arbitrary scripts. No file transfer to a device. If something is not on this list, the answer is that a person does it with their own hands and records what they did.

##The audit log

Every approval, every action, every declined suggestion and every permission change is recorded with the actor, the time and the device state either side. The log is append-only and visible to admins in full.

Nobody can delete an entry, including us. Correcting a mistake means adding a record, not removing one.

audit · recent

18 Sep 09:21  ravi.k      approved  service.restart · desk-04 · spooler
18 Sep 09:34  ravi.k      declined  patch KB5062553 · sensitive ring
                                    reason: month-end, revisit 02 Oct
18 Sep 11:02  meena.r     changed   priority #2240 low → medium
18 Sep 18:04  ravi.k      approved  identity.revoke · arjun.m · all sessions
19 Sep 09:12  system      closed    leaver checklist · arjun.m

##Data handling

In transit: TLS 1.3, certificate pinned in the agent. An intercepting proxy will stop the agent reporting rather than let it report through.
At rest: AES-256 on the database and on backups.
Region: chosen at workspace creation and fixed after it. Data does not move between regions, including for support.
Export: everything, at any time, as CSV and JSON - devices, tickets with full history, licences, patch records, the audit log.
Retention: on cancellation the workspace is readable for 30 days, then deleted - backups included - within a further 30.
Support access: our staff cannot read your workspace. Access needs a written request from an admin, is scoped to a time window, and appears in your audit log as it happens.

##Who inside your workspace can see what

Role	Can see	Can approve
Requester	Their own tickets and their own assigned devices	Nothing
Operator	All tickets, all devices, licences, patch state	Only what an admin has delegated, by category
Admin	Everything, including the full audit log	All actions
Client read-only	Their own org's tickets and device summary	Nothing

##Limits

Limits - what this does not do

This is not endpoint protection. It inventories your antivirus and reports its state; it does not detect or block threats.
Not a remote access tool. There is no screen sharing and no remote shell, so an incident still needs your own remote support tooling.
Not a backup product. It records whether a backup agent is installed and reporting, not whether your data is recoverable.
Not a DLP or monitoring product. It cannot tell you what a person did, by design, and no setting changes that.
Certificate pinning means TLS-inspecting middleboxes will break agent reporting. Allow the agent host to pass through unaltered.