MITRE ATT&CK Enterprise · TA0007
Discovery: the intruder maps your estate before touching it
DiscoveryTA0007
Before anyone steals anything, they have to find out what there is to steal. This is that phase — what it looks like from the outside, and the telemetry that catches it.
For the owner
An intruder spends their first hours doing nothing you would notice.
They are not breaking in, not stealing, not asking for ransom. They are reading: the list of people who work here, the machines on the office network, the shared folders, the group memberships, the security software that is running. Then they pick one machine and move.
That quiet phase is called discovery, and it is the reason a breach can look harmless for days. From the outside nothing happens: no system is down, no customer complains, no money moves. The cost arrives later, and it is always bigger because of what the intruder learned — the payroll export they found, the backup server they located, the account they could reuse without ever needing a password.
Three questions are worth asking your team, and none of them needs new spending to answer:
- Do we actually know how many machines, accounts and shared folders we have? Most discovery succeeds because the inventory an attacker reads is more complete than the inventory the business believes in.
- Could somebody read our staff list for a week without touching anything, and would we know? Enumeration that stays inside ordinary administration looks exactly like administration.
- How long would it take an outsider to map our estate? If the answer is "a few hours", then the map is the real cost of the attack, and every later step is cheaper because of it.
The reason this is worth making visible is that discovery is where a small intrusion becomes a company-wide one. The intruder who understands the layout chooses better targets than the intruder who does not, and the gap between those two intruders is measured in days of quiet, uneventful reading.
For the analyst
The tactic
MITRE ATT&CK files Discovery under TA0007. It is the tactic that answers a single question — what is here? — and the matrix files 49 techniques under it, while the pinned SigmaHQ corpus carries 219 rules that name a technique of this tactic.
Discovery is the quietest tactic in the matrix and also the widest, and that is not a contradiction. An intruder who wants to move through an estate has to answer several separate questions, and ATT&CK gives each of them a name. The number of techniques under this tactic measures how many things there are to learn, not how many attacks there are.
| Technique | Name | Rules in the corpus |
|---|---|---|
T1082 | System Information Discovery | 31 |
T1033 | System Owner/User Discovery | 29 |
T1083 | File and Directory Discovery | 22 |
T1087.002 | Account Discovery: Domain Account | 20 |
T1046 | Network Service Discovery | 19 |
T1482 | Domain Trust Discovery | 17 |
T1018 | Remote System Discovery | 16 |
T1069.001 | Permission Groups Discovery: Local Groups | 15 |
T1087 | Account Discovery | 15 |
T1069.002 | Permission Groups Discovery: Domain Groups | 14 |
T1087.001 | Account Discovery: Local Account | 12 |
T1012 | Query Registry | 9 |
T1016 | System Network Configuration Discovery | 9 |
T1040 | Network Sniffing | 9 |
T1007 | System Service Discovery | 8 |
T1518.001 | Software Discovery: Security Software Discovery | 8 |
T1049 | System Network Connections Discovery | 7 |
T1057 | Process Discovery | 6 |
T1201 | Password Policy Discovery | 5 |
T1615 | Group Policy Discovery | 5 |
T1135 | Network Share Discovery | 4 |
T1217 | Browser Information Discovery | 4 |
T1518 | Software Discovery | 4 |
T1069 | Permission Groups Discovery | 3 |
T1087.004 | Account Discovery: Cloud Account | 3 |
T1124 | System Time Discovery | 3 |
T1497.001 | Virtualization/Sandbox Evasion: System Checks | 3 |
T1526 | Cloud Service Discovery | 3 |
T1120 | Peripheral Device Discovery | 2 |
T1614.001 | System Location Discovery: System Language Discovery | 2 |
T1010 | Application Window Discovery | 1 |
T1069.003 | Permission Groups Discovery: Cloud Groups | 1 |
T1580 | Cloud Infrastructure Discovery | 1 |
T1613 | Container and Resource Discovery | 1 |
T1619 | Cloud Storage Object Discovery | 1 |
T1622 | Debugger Evasion | 1 |
The four questions discovery asks
"What machine am I on, and who owns it?" — System Information Discovery (T1082) and System Owner/User Discovery (T1033). The first thing any impostor does is establish their own footing: the host name, the operating system and its patch level, the architecture, the domain membership, the time zone. On Linux the same answer arrives as uname -a, cat /etc/os-release, id and hostname; on Windows it is systeminfo, wmic, whoami /all. T1082 is the largest block of rules under the tactic and the least alarming in isolation, because every administrator runs some version of it.
The reason it is still worth a rule is the context: T1033 collects the account the process runs as, and that account is the interesting half. The normalizer puts it in user_name, the utility in process_name, its parent in parent_process, and the whole invocation in command_line. A whoami run by a workstation account on a workstation is noise. The same whoami run by a service account on a database server is not.
"Who else exists, and what are they allowed to do?" — Account Discovery (T1087, T1087.001 local, T1087.002 domain, T1087.004 cloud) and Permission Groups Discovery (T1069, T1069.001 local groups, T1069.002 domain groups, T1069.003 cloud groups). This is the most valuable discovery in the tactic, because the account list is also the attacker's shopping list. The mechanics are well known: net user /domain, net localgroup administrators, net group "Domain Admins" /domain, dsquery, Get-ADUser, Get-ADGroup, and on Linux getent passwd and getent group.
The telemetry is process creation plus identity. A net.exe or net1.exe spawned from cmd.exe with /domain in the command line is a Windows event; keycloak emits an administrative event when a realm administrator lists users; m365_audit and gworkspace record the same thing in the cloud, where there is no local process at all. The normalized columns make those three sources comparable, which is the point: the same question, asked of a Windows directory and of a cloud tenant, produces the same shape of row.
"What is stored here?" — File and Directory Discovery (T1083), Network Share Discovery (T1135), Software Discovery (T1518) with its Security Software Discovery sub-technique (T1518.001), Query Registry (T1012), and Group Policy Discovery (T1615). dir /s, net view, wmic product get, reg query, gpresult /r, Get-MpComputerStatus. T1083 is the third-largest block of rules under the tactic.
T1518.001 deserves its own attention: the intruder is not asking what is installed, they are asking what is watching them. The registry query under HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\Uninstall and the security centre namespace are the same enumeration that a patch audit performs. If the answer is "an endpoint agent is running", everything the intruder does afterwards gets shaped by it.
"What am I connected to?" — Network Service Discovery (T1046), Remote System Discovery (T1018), System Network Configuration Discovery (T1016), System Network Connections Discovery (T1049), Network Sniffing (T1040) and Domain Trust Discovery (T1482). netstat -ano, arp -a, ipconfig /all, route print, nltest /domain_trusts, and a raw socket put into promiscuous mode for the sniffing variant. Here the endpoint telemetry and the network telemetry meet: the enumeration is a process, but the answer is a set of dst_ip and dst_port pairs that the firewall logs also carry, and firewall_syslog shows the connections the intruder was looking for.
The rest of the tactic is discovery with a narrower target: T1057 process discovery, T1007 service discovery, T1201 password policy discovery, T1124 system time discovery, T1120 peripheral device discovery, T1614.001 system language discovery, T1010 application window discovery, and T1217 browser information discovery, where the target is the browser's history and extensions rather than the machine's. Two families point outward instead of inward: T1497.001 checks the machine for virtual-machine artefacts, which is discovery aimed at a sandbox, and the cloud family — T1526 cloud service discovery, T1580 cloud infrastructure discovery, T1613 container and resource discovery, T1619 cloud storage object discovery — asks the same four questions of a tenant instead of a domain.
Where the telemetry comes from
For this tactic the sources divide cleanly by who does the asking:
- Endpoint process creation (
windows_sysmon,linux_auditd). The enumeration utility, its arguments and its parent. The highest-value source, because the command line is where the intent lives. - Identity and directory logs (
keycloak,m365_audit,gworkspace,cloud). Administrative reads of the user list, group list and role assignments. No process here — the enumeration happened in the directory. - Network logs (
firewall_syslog). The topology the intruder was trying to map, recorded assrc_iptodst_ipwith adst_port. - File and path activity. The
file_pathof what was read, theuser_namethat read it, theprocess_namethat opened it.
How a rule actually reasons over this
Discovery is high-volume benign activity, so a rule that fires on the event itself is guaranteed to be muted. The rarity has to come from a conjunction. These are the anchors that work:
| Signal | Column | Why it is rare |
|---|---|---|
| Enumeration by a non-administrator | user_name, process_name | Admin tools on an admin account are the baseline |
| Enumeration from a host that never enumerated before | host_id, process_name | The account is the constant; the host is not |
| Enumeration outside working hours | event_ts, user_name | Identity logs carry no clock skew; endpoint clocks do |
| Discovery of security software by a process that is not a package manager | process_name, command_line | The legitimate path is a patch or inventory run |
| Directory enumeration of cloud storage | source_type, file_path | Cloud enumeration comes from an audit trail, not a process |
A rule therefore looks like a rarity test rather than a name test:
title: Domain enumeration from an account that does not administer the domain
status: experimental
logsource:
category: process_creation
product: windows
detection:
parent: 'cmd.exe|powershell.exe|pwsh.exe'
enumeration: 'net.exe|net1.exe|dsquery.exe|whoami.exe'
admin_group: 'Domain Admins|Enterprise Admins'
filter_known_admin_tools:
user_name|contains: 'svc_'
parent|contains: 'admin'
condition: parent and enumeration and not filter_known_admin_tools
fields: [user_name, host_id, process_name, parent_process, command_line, src_ip]
Two things in that sketch are worth stating outright. The filter is not decoration: without it the rule fires on the help desk's own enumeration and gets switched off within a week. And the user_name is the discriminator, not the binary — net.exe is the same binary in both cases.
Where analysts get this wrong
The classic mistake is treating discovery as a noise problem and solving it by narrowing the rules until they are quiet. That trades a false positive for a false negative, and the false negative here is expensive: a single account that enumerated the domain on one afternoon and then used what it found is invisible for the whole intrusion.
The second mistake is under-counting the sources. Discovery happens at least as much in the directory and the cloud as on the endpoint. A deployment that collects only endpoint telemetry sees an intruder enumerate the local machine, read a few directories, and disappear — because everything interesting was happening in a query that never touched the endpoint. The identity and network streams are not a supplement to discovery coverage; they are half of it.
The third is confusing enumeration with usage. T1087 asks who exists; credential access asks what they can use. The difference between the two is rarely the command — it is the account, the host and the hour.
Detection engineering notes
- Anchor on the conjunction, never the tool. Every mainstream enumeration utility is dual-use and often installed. What makes the event hostile is the account, the host and the hour together.
- Count the rare thing, not the event.
count()byuser_nameover a window turns "someone listed the domain" into "this account listed the domain eleven times in ten minutes", which is a different alert with a different urgency. - Rate-limit per account, not per event. An account that enumerates constantly generates more events than anyone will read. Alerting once per account per day keeps the signal and loses nothing.
- Write against the normalized columns.
CommandLineandcommand_line,Imageandprocess_name— the corpus is not consistent, the normalizer is. A rule written against raw collector names works on exactly one host. - Know the baseline before the rule. Which accounts legitimately enumerate the domain, and from which hosts, is the difference between a rule that fires and a rule that is muted on day one.
Where DetectMonkey fits
Discovery is, before anything else, an accounting problem. The telemetry an intruder uses to enumerate an estate is the same telemetry an administrator uses every day, so the useful question is not "was this hostile?" but "how much did this account learn, and does that amount match the job?"
DetectMonkey answers that by keeping the discovery-shaped events — process creation, directory listings, identity queries, share and service scans — in one normalized shape, so a week of enumeration becomes a set of rows an analyst can filter by account, by host and by role instead of a queue of alerts to read one at a time. The pinned corpus of 3164 Sigma rules evaluates them as the events arrive, so an enumeration burst is visible while it is still running, and the last 72 hours stay queryable afterwards to reconstruct exactly what was learned.
The platform detects and alerts. It does not isolate, quarantine or block anything on your systems, and it does not change your configuration. The telemetry is processed and stored on private servers operated by Codein in New Brunswick, Canada, not in a public cloud, and analysts work in English and French. Early access is 90 days at no charge, for up to 100 seats, founder-led onboarding and no card required.
See it on your own telemetry
Early access is 90 days at no charge, for up to 100 seats, founder-led onboarding, no card required.