MITRE ATT&CK Enterprise · TA0009
Collection: the attacker gathers what they came for
CollectionTA0009
The break-in is over and the noise has stopped. Now the intruder is reading your files, your mail and your screens and packing them for the trip out. This is the ATT&CK tactic that decides what a breach is actually worth.
For the owner
The noiseless part of an intrusion is the part that decides what it was worth.
An intruder who has broken in does not stay there to break things. They read. They copy the payroll export, the client list, the medical records, the design documents, the board minutes. They gather it quietly, they pack it, and they take it with them. Nothing breaks. No system is down. From the outside, the day looks entirely normal.
That is what makes collection the phase an owner should worry about most, and the phase that is hardest to notice. A copy leaves no mark on the original: the file you are missing is still sitting on the server, with the same modification time and the same owner. The evidence is not in what disappeared, it is in what was read, by whom, and where it went next.
And what was collected decides everything that follows. The difference between an incident and a catastrophe is often one file: the one that holds customer personal information, the one that holds the credentials, the one that a contract says must never leave the building.
Three questions are worth asking your team:
- What is the most sensitive thing we store, and who can read it? If the answer is "several folders and a shared drive", that is your exposure, and it is much larger than the number of systems you own.
- If somebody quietly copied a whole folder this week, would we know? The honest answer is usually no, because nothing changed on the source.
- Which of our systems hold obligations we owe to other people? Client files, health records, employee data — these turn a security incident into a reporting obligation, and the reporting obligation is what regulators and insurers measure.
The reason to make collection visible is simple. By the time you notice a breach, the collection has usually been finished for days, and the only thing left to decide is how much you tell whom.
For the analyst
The tactic
MITRE ATT&CK files Collection under TA0009. It is the tactic that answers what did they take?, and the matrix files 41 techniques under it while the pinned SigmaHQ corpus carries 107 rules that name a technique of this tactic.
Collection is the tactic analysts most often under-detect, for a structural reason: it leaves almost no trace where the analyst is looking. A read does not modify a file. A copy does not change the source. The file-integrity check is silent, the backup still succeeds, and the user is still logged in. Everything that proves the collection happened is somewhere else — in the process that read the file, in the archive that was created, in the volume that left, in the audit trail of a mailbox.
| Technique | Name | Rules in the corpus |
|---|---|---|
T1560.001 | Archive Collected Data: Archive via Utility | 13 |
T1005 | Data from Local System | 12 |
T1557.001 | Adversary-in-the-Middle: Name Resolution Poisoning and SMB Relay | 10 |
T1113 | Screen Capture | 9 |
T1557 | Adversary-in-the-Middle | 9 |
T1115 | Clipboard Data | 7 |
T1213 | Data from Information Repositories | 7 |
T1123 | Audio Capture | 6 |
T1213.003 | Data from Information Repositories: Code Repositories | 5 |
T1074.001 | Data Staged: Local Data Staging | 4 |
T1114 | Email Collection | 4 |
T1119 | Automated Collection | 4 |
T1056.001 | Input Capture: Keylogging | 3 |
T1056.002 | Input Capture: GUI Input Capture | 3 |
T1039 | Data from Network Shared Drive | 2 |
T1056 | Input Capture | 2 |
T1074 | Data Staged | 2 |
T1114.003 | Email Collection: Email Forwarding Rule | 2 |
T1185 | Browser Session Hijacking | 2 |
T1560 | Archive Collected Data | 2 |
T1114.001 | Email Collection: Local Email Collection | 1 |
T1125 | Video Capture | 1 |
T1557.003 | Adversary-in-the-Middle: DHCP Spoofing | 1 |
Gathering
Data from Local System (T1005) is the base case: the local disk, the user's documents and downloads, the browser profile, the cached mail store. Data from Information Repositories (T1213) widens it to the systems people actually keep their work in — the source repository (T1213.003), the wiki, the document library, the shared drive (T1039), the ticketing system. In a modern tenant these are cloud products, and the telemetry is an audit trail rather than a process: m365_audit records file access, download and sync activity; gworkspace records the Drive and Docs equivalents.
Email Collection (T1114) comes in three forms, not equally hard to see. Local collection (T1114.001) is the mail client on the workstation. The interesting one is T1114.003, the forwarding rule: a rule created inside the mailbox that sends copies of incoming mail to an address outside the tenant. It is persistent, it survives a password change, and it lives in the mailbox configuration rather than in a log file, which makes it both easy to create and easy to miss.
Clipboard Data (T1115) is a small thing that carries large things: a password pasted from the password manager, a contract copied out of a web page, a client list from a spreadsheet. Detection means watching for processes that call the clipboard API — a script that reads the clipboard outright, or a collection tool doing it in a loop.
Screen Capture (T1113) and its relatives — Audio Capture (T1123), Video Capture (T1125), and Input Capture (T1056), whose sub-techniques cover keylogging (T1056.001) and capturing what the user types into a window (T1056.002) — collect what the user sees and does rather than what is stored. The process is the signal: a screen capture is usually an interpreter calling the graphics API, and the call itself is in the command_line.
Browser Session Hijacking (T1185) takes the session cookie instead of the credential, which means the intruder inherits a session without ever seeing a password.
Automated Collection (T1119) is the pattern rather than the source: the same collection script, run across many hosts, as one account. The tell is repetition with the same shape on different host_id values.
Adversary-in-the-Middle
Adversary-in-the-Middle (T1557) collects data by intercepting it in transit. The name resolution and SMB relay variant (T1557.001) sits between a client and a file server and relays the traffic, so the file is captured while the client believes it is talking to the server. The DHCP variant (T1557.003) points the victim at the attacker's host, and the victim's credentials arrive with the first request.
Staging and archiving
This is where collection becomes visible, and it is the most productive half of the tactic for detection.
Data Staged (T1074) and its local variant (T1074.001) are the intruder collecting loose ends into one place before packing. The staging locations are well known and worth writing into a rule by hand: the temporary directory, C:\ProgramData, C:\Users\Public, the Windows temp directory, and the equivalent paths on Linux. The point is the pattern — several distinct source files ending up in one directory.
Archive Collected Data (T1560) and Archive via Utility (T1560.001) are the packing. A 7-Zip, WinRAR, tar, zip or gzip process started with an output path, and the resulting artefact written to disk. This is the strongest single signal in the tactic, because the archiver is a process creation with a name, a parent and a command line, and because the artefact is a file write with a path.
The reason it is also the most fragile is worth stating: the utility list is long and environment-specific. 7z.exe, 7za.exe, 7zr.exe, WinRAR.exe, rar.exe, tar.exe, gzip, zip.exe, cab.exe, bsdtar — and the same utility may be named differently in the container, on the Linux host, or in the build environment. A rule that lists one binary sees one binary.
Where the telemetry comes from
- Endpoint process creation and file activity (
windows_sysmon,linux_auditd). The archiver and its arguments inprocess_nameandcommand_line, the artefact infile_path, the account inuser_name. - Cloud and identity audit (
m365_audit,gworkspace,cloud). File access and download events, mail items accessed, and — the valuable one — a forwarding rule created on a mailbox. - Network logs (
firewall_syslog). The bulk of the collection leaving for a destination the account has never used, and the relay traffic of an interception. - Session and token records (
keycloak). The account whose session suddenly covers a hundred file paths in an hour.
How a rule actually reasons over this
Collection events are cheap and plentiful, so the rule has to reach for context every time. These are the anchors that hold:
| Signal | Columns | Why it is rare |
|---|---|---|
| Archiver run by a non-administrator, or on a host with no archive history | process_name, user_name, host_id | Backups and packaging are the baseline; an archiver is not |
| Archive written outside the user's own documents | file_path, user_name | A user archives into their own folder |
| Mailbox forwarding rule created | source_type, user_name | Almost no mailbox has ever had one |
| Screen capture on a server | process_name, host_id | Servers have no screens |
| One account reading hundreds of paths in an hour | user_name, file_path | People read dozens of files, not hundreds |
| Collection by the same shape of process across many hosts | host_id, process_name | That is T1119, and it is machine-like |
A rule for the packing stage looks like this:
title: Archive utility invoked to package data outside the user workspace
status: experimental
logsource:
category: process_creation
product: windows
detection:
archiver: '7z.exe|7za.exe|7zr.exe|WinRAR.exe|rar.exe|tar.exe|zip.exe'
staging_path: '\\Temp\\|\\ProgramData\\|\\Users\\Public\\'
selection:
- 'file_path|contains|':
- 'command_line|contains|':
filter_backup:
process_name|contains: 'backup'
condition: archiver and staging_path and not filter_backup
fields: [user_name, host_id, process_name, parent_process, command_line, file_path]
Three things in that sketch are engineering, not decoration. The archiver list has to be long, or the rule misses. The staging path separates packing from the legitimate compression of a user's own folder. And the archive's size — which arrives in the event payload rather than in a column of its own — turns "someone compressed a folder" into "someone compressed something they should not have had".
The better rules correlate backwards rather than forwards: the archiver is interesting because of the reads that preceded it. A window of file access by one account, followed by one archive, beats either half.
Where analysts get this wrong
The first mistake is looking for the wrong trace. The instinct is to watch for files being changed or deleted, because that is what intrusion detection usually means. Collection does neither: the file is still there afterwards with its original timestamp. Everything that matters about the collection is in the process, the account, the volume and the destination — never in the state of the source file.
The second mistake is under-reading the cloud. In a tenant where the work lives in Drive, SharePoint, OneDrive or a mail system, the collection may never touch an endpoint: it is an API call, and the only record is the audit trail. A deployment with endpoint telemetry and no cloud audit sees an intruder sit still for a week and take everything.
The third is confusing collection with legitimate heavy usage. A backup agent, a sync client, a migration script and a person working through a shared drive all look like bulk access. The differentiator is not the volume, it is the account and the destination: where did the data go, and was it ever going to go there?
Detection engineering notes
- Watch the second stage, not the first. The archive is a discrete event with a name, a parent and a path; the reads before it are a blur.
- Never rely on the source file. A rule that reasons about modification or deletion will not fire.
- Cover the archive utility list in full, per environment. Enumerate the archivers actually installed on your estate, including the ones nobody admits to.
- Treat the mailbox as a collection target. A forwarding rule is created once and then works forever; that asymmetry is what makes it worth a rule.
- Correlate reads with the archive. The window between the reads and the packing is the window in which the intrusion is still recoverable.
- Use the account, not the file.
user_nameseparates the user from the intruder wearing the user's credentials.
Where DetectMonkey fits
The 72 hours after a collection event are the most useful hours of the whole incident, and they are the hours most companies spend empty.
DetectMonkey keeps that window populated. The collectors read the telemetry a collection leaves behind — the archive utility and its arguments, the files read and by whom, the mail rules created, the bulk access and the destination it went to — and the pinned corpus of 3164 Sigma rules evaluates them as the events arrive, so the packing is seen while it happens rather than reconstructed afterwards. Because the last 72 hours stay queryable and the cold lake keeps 365 days, the question "what did this account read, and where did it go?" is answerable with one query, for an incident that started three days ago or three months ago.
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.