MITRE ATT&CK Enterprise · TA0010
Exfiltration: how the data leaves
ExfiltrationTA0010
The break-in is the part everyone worries about. The copy is the part that costs you. This is what data leaving looks like, what it costs, and the telemetry that catches it.
For the owner
The worst day of an intrusion is not the day the intruder got in. It is the day they left with something.
By the time a copy of your customer list, your payroll file or your patient records turns up in someone else's hands, the intrusion is already over. So the question for an owner is not "can we stop this" — it is "would we know, and how fast".
Data leaves in ways that look boring. A folder uploaded to a personal cloud account. A database copied into a file and sent over an ordinary web connection. A share copied to an external drive. To a network device, each of those is normal work, and each of them is invisible to a system that only watches for malware.
The useful signal is never "data left". It is "data left in a way this machine has never done it before": a first-time destination, a first-time tool, an unusual hour, a volume that does not match the person's job.
Three questions are worth asking your own team:
- Which people can export data, and can you tell afterwards what they took?
- Do you have a record of outbound traffic, or only of inbound alarms?
- If a copy of your data were taken tomorrow, how long until someone on your team would know?
None of that needs a bigger budget. It needs the trail to exist, and someone reading it.
For the analyst
The tactic
MITRE ATT&CK files Exfiltration under TA0010. It answers one question: by which door did the data leave? The matrix files 20 techniques under it, and the pinned corpus carries 59 rules that name a technique in this tactic.
Exfiltration is unusual because the answer is often mundane. ATT&CK splits it by the channel the data travelled through, not by the value of what was taken: the same theft looks different over a consumer cloud product, a DNS tunnel, or the command channel the attacker already owned.
| Technique | Name | Rules in the corpus |
|---|---|---|
T1567.002 | Exfiltration Over Web Service: Exfiltration to Cloud Storage | 13 |
T1567 | Exfiltration Over Web Service | 11 |
T1048 | Exfiltration Over Alternative Protocol | 9 |
T1020 | Automated Exfiltration | 8 |
T1048.003 | Exfiltration Over Alternative Protocol: Exfiltration Over Unencrypted Non-C2 Protocol | 8 |
T1537 | Transfer Data to Cloud Account | 6 |
T1030 | Data Transfer Size Limits | 2 |
T1041 | Exfiltration Over C2 Channel | 2 |
T1567.001 | Exfiltration Over Web Service: Exfiltration to Code Repository | 2 |
T1048.001 | Exfiltration Over Alternative Protocol: Exfiltration Over Symmetric Encrypted Non-C2 Protocol | 1 |
The three doors
Over a web service (T1567, T1567.001, T1567.002, T1537). The attacker uploads stolen material to a service that already looks like normal business traffic to the network: a cloud storage provider, a code hosting platform, a file-sharing product. The corpus reflects exactly this. There are rules on DNS queries for consumer file-sharing domains, on network connections to the Dropbox API domain, on the user agent string of rclone seen in a web proxy log, and on the appearance of an rclone configuration file on a workstation — the profile that says where the uploads go.
The cloud audit trail carries the same behaviour from the other side. A Microsoft 365 rule fires on data exfiltration to an unsanctioned application; GitHub audit rules fire on a private repository being forked, transferred or switched to public. T1567.001 is the specific case of code: a source tree pushed to a repository, or a tunnel endpoint published to a service that resolves by name.
Over an alternative protocol (T1048, T1048.001, T1048.003, T1041). The attacker uses a protocol that is not the command channel, so traffic inspection that only watches the command channel sees nothing. In the corpus this shows up as ICMP and DNS-based exfiltration from PowerShell, a WebDAV PUT request seen by a network sensor, wget and curl file upload on Linux, a Python HTTP server started where no HTTP server belongs, and rundll32 being used as a WebDAV client. T1048.003 is the unencrypted variant and T1048.001 the encrypted one; the split matters because the encrypted form defeats content inspection.
T1041 is the other end of the same idea: stealing data through the command-and-control channel the attacker already owns. There is no new destination to learn, no new tool to notice — only the volume.
Automated and throttled (T1020, T1030). Loss of data during a period of time, rather than a single transfer. The classic technique is to wait for a quiet period and take everything at once; the classic detection is a rule that watches for a file being split into pieces, on Linux and on macOS. The throttling variant matters more than it looks: an attacker who transfers the same volume in small pieces over a week stays under every volume threshold anyone has ever written.
Where the telemetry comes from
Exfiltration is the tactic where the endpoint is least useful and the network is most useful, because the data has to leave the machine to go anywhere.
- DNS query events. The cheapest signal and often the earliest. A workstation that has never resolved
mega.nzand then resolves it a hundred times is not doing anything a finance analyst does. Consumer file-sharing and paste sites are named in the corpus for this reason. - Network connection events. The destination address and port, and the process that opened them. A connection to a cloud storage API from a process that has never spoken to that API is worth an alert.
- Web proxy logs. Where the organisation has them, the user agent and the request path expose the tooling (
rclone,curl,wget) and the destination, including services that are otherwise indistinguishable from ordinary browsing. - Firewall and perimeter syslog. The only source that sees the transfer when the endpoint is silent — for example a server that was compromised before the collector was installed.
- Process creation and file events. The staging side:
curlwith an upload flag, awgetto a remote host, the creation of anrcloneconfiguration file, the archive utility that compressed the staging directory before the transfer. - Cloud audit logs. Where the copy lands: an application consent granted, a bulk download, a repository made public.
How a rule actually reasons over this
Most of the corpus does not detect "a lot of data left". It detects a destination the host has never used before, and it needs a per-host baseline to do that. Two events carry that state: the DNS query and the outbound connection. Either one alone is weak; together they are strong, and the DNS query usually arrives first, seconds before the connection opens.
The reasoning that makes a rule good here has three parts:
- Rare destination. A new domain for this host, resolved for the first time, by a process whose normal behaviour does not involve it.
- Rare volume. A transfer size or a request count above what this account, on this machine, has ever produced.
- Rare tool.
rclone,7z,megatools, acurlupload, awgetto an IP address instead of a name. The corpus encodes exactly these strings.
What makes it wrong is the ordinary overlap. Backup software that uploads to cloud storage, an editor that uploads to a file-sharing service, a developer pushing a build artefact to a repository — all of them look like theft. Those rules are where false positives live, and where a per-tenant allowlist pays for itself.
Detection engineering notes
- Correlate the read with the transfer, not the transfer alone. A rule that fires on "large outbound transfer" is a volume alarm, and volume alarms are the first thing a team turns off. A rule that fires on "a sensitive file was read and a new destination was contacted within the same window" is a detection.
- Keep the chunking rules. The corpus rule for a file being split into pieces is worth far more than a hundred volume thresholds, because splitting is only done when you intend to defeat exactly those thresholds.
- Watch the code path, not just the file path.
T1567.001is rare and cheap: a repository that goes public, a dev tunnel that publishes a local port. These are configuration events with excellent signal. - Treat a sanctioned cloud as a suspect, not as safe. The interesting case is the copy that goes to a cloud account you did not sanction, not the one that goes to the drive your company already paid for.
- Know the blind spot. If the endpoint collector was installed after the compromise, the network still sees the transfer and the endpoint does not. Rules that need
process_namewill not fire; rules that need onlydst_ipwill.
Where analysts get this wrong
The most common error is timing. Teams build excellent exfiltration rules and then discover that the alert arrives after the data is gone, because the collection tactic fired days earlier and the exfiltration alert fires once the transfer completes. The two have to be correlated, in both directions.
The second error is assuming the destination has to be new. An attacker who knows your company uses a cloud storage provider does not use a different one. They use the same one, with an account they created, which means the only unusual thing is the account, the volume and the hour. T1537 exists in the matrix for exactly this: data moved to an adversary-controlled cloud account.
The third error is treating exfiltration as terminal. It usually is followed by impact — encryption, or a threat to publish — and the most valuable work is the correlation that connects the copy to the extortion.
Where DetectMonkey fits
Detection of exfiltration is an analytics problem before it is a sensor problem. Data has to go somewhere; the work is recognising that "somewhere" as unfamiliar, and recognising it while the transfer is still small enough to matter.
DetectMonkey watches for exactly that. It ingests the DNS, connection, proxy and file events this tactic produces, evaluates them against the pinned corpus of 3164 rules as the events arrive rather than after they are indexed, and keeps the last 72 hours queryable so an analyst can reconstruct the route the data took, not merely the fact that it left.
The platform detects and alerts. It does not isolate, quarantine or block anything on your systems, and it changes no configuration of yours. The telemetry is processed and stored on private servers operated by Codein in New Brunswick, Canada. Every alert carries its evidence in English and French.
Early access is 90 days at no charge, for up to 100 seats, with 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.