MITRE ATT&CK Enterprise · TA0011
Command and Control: the channel the intruder keeps open
Command and ControlTA0011
An intruder who cannot call home cannot be told what to do next. This is the ATT&CK tactic for the channel back — the beacon, the proxy, the fallback — and the telemetry that shows it.
For the owner
An intruder who cannot call home is an intruder who cannot finish anything.
Somewhere between the first machine they broke into and the last thing they take, there is a channel. It is how they fetch their tools, learn what they have caught, receive the instruction to encrypt, and know when to stop. An attacker with a working channel can turn one compromised workstation into a campaign across your estate. An attacker without one is stuck with what they found on the first day.
From the outside, a channel is almost invisible, and that is exactly why it is worth watching. A computer talking to the internet is the normal condition of every computer you own. There is no outage to report, no customer to call, no ransom note. The only thing that changes is that one machine in the building now has a regular, slightly odd rhythm of outbound traffic, and one device on your list is doing something no other device on your list does.
The business consequence is about control. A channel means the intruder keeps decisions for later: what to encrypt, which server to reach next, when to stop. It also means the attack is not finished when you think it is. Six weeks after the break-in, a machine can still be taking instructions, and the question "is this still infected?" cannot be answered by asking the machine.
Three questions are worth asking your team:
- If one device on our list started talking to somewhere new every hour, would anyone notice? Most estates cannot answer this today, and the answer is usually no.
- Which of our devices are allowed to reach the internet, and which destinations are they allowed to reach? A list of approved destinations is the cheapest version of this control and the easiest to start.
- How would we know a machine was still compromised if the intruder simply stopped? The channel is how a compromise stays alive; it is also how it stays quiet.
The reason to make the channel visible is that everything else in an intrusion is downstream of it. The tools, the instructions and the timing all arrive through it, and it is the one place where a live conversation can be observed rather than inferred.
For the analyst
The tactic
MITRE ATT&CK files Command and Control under TA0011. It is the tactic that describes the adversary's channel back, and the matrix files 45 techniques under it while the pinned SigmaHQ corpus carries 233 rules that name a technique of this tactic.
It is a large tactic, and deliberately so: there are many ways to build a channel and many ways to hide one. What the sub-techniques share is not the tool but the shape of the traffic — an application protocol, a non-standard port, a proxy, a tunnel, a service the attacker controls, a domain that changes every day.
| Technique | Name | Rules in the corpus |
|---|---|---|
T1105 | Ingress Tool Transfer | 70 |
T1219.002 | Remote Access Tools: Remote Desktop Software | 43 |
T1071.001 | Application Layer Protocol: Web Protocols | 31 |
T1572 | Protocol Tunneling | 23 |
T1090 | Proxy | 21 |
T1071.004 | Application Layer Protocol: DNS | 13 |
T1102 | Web Service | 13 |
T1090.001 | Proxy: Internal Proxy | 6 |
T1219 | Remote Access Tools | 6 |
T1071 | Application Layer Protocol | 5 |
T1571 | Non-Standard Port | 5 |
T1008 | Fallback Channels | 4 |
T1102.001 | Web Service: Dead Drop Resolver | 4 |
T1132.001 | Data Encoding: Standard Encoding | 4 |
T1573 | Encrypted Channel | 4 |
T1090.003 | Proxy: Multi-hop Proxy | 3 |
T1095 | Non-Application Layer Protocol | 3 |
T1102.002 | Web Service: Bidirectional Communication | 3 |
T1568 | Dynamic Resolution | 3 |
T1001.003 | Data Obfuscation: Protocol or Service Impersonation | 2 |
T1090.002 | Proxy: External Proxy | 2 |
T1102.003 | Web Service: One-Way Communication | 2 |
T1568.002 | Dynamic Resolution: Domain Generation Algorithms | 2 |
The channel families
Application Layer Protocol (T1071) is the base case: the intruder speaks a protocol the network is meant to allow, so the traffic looks ordinary. The sub-technique carrying most of the rules is T1071.001, web protocols — HTTP and HTTPS to a server the attacker controls, with the payload in the request, the response or a cookie. It is the hardest family to detect by content, because a request to a web server is what a web server is for. What gives it away is the combination: a process with no business on the network making an outbound HTTPS connection to a destination never seen on that host.
T1071.004 is the DNS variant, and the one that still works when egress is filtered. The payload is encoded into DNS queries — T1132.001, standard encoding, is the matching technique for the encoding step — so the channel rides on name resolution traffic. The signal is the shape of the lookups: a domain with high entropy, a very long label, a request type that does not match the host's business, or a burst of NXDOMAIN answers from one machine.
Web Service (T1102) turns a legitimate service into the channel: the adversary registers a domain and uses it as a rendezvous point, in one of three arrangements — T1102.001, a dead drop resolver, where the victim fetches instructions from a public page; T1102.002, bidirectional communication, where the same service carries traffic both ways; and T1102.003, one-way communication, where the victim only sends. The dead drop is the worst case for a detection programme: it produces no inbound connection at all, only an outbound fetch of a page that looks like a page.
Non-Application Layer Protocol (T1095), Non-Standard Port (T1571) and Protocol Tunneling (T1572) cover the cases where the channel does not look like an application at all. A raw socket on a port nothing in the estate uses is T1571; a custom protocol with its own framing is T1095; wrapping the traffic inside another protocol — DNS inside HTTPS, SSH inside HTTP, a tunnel client — is T1572, and it defeats naive allowlists, because the destination looks legitimate and the port is the expected one.
Proxy (T1090) is the intruder borrowing someone else's network position, in three shapes: an internal proxy (T1090.001) inside the compromised estate, so the traffic appears to come from inside; an external proxy (T1090.002) rented or abused, so it appears to come from elsewhere; and a multi-hop proxy (T1090.003), where the chain is long enough that the origin is genuinely hard to establish.
Encrypted Channel (T1573) and Data Obfuscation (T1001.003) are the opposite problem: the channel is visible but unreadable, and the traffic impersonates a service it is not. This is where the honest limits of detection have to be stated: with an encrypted payload on a well-behaved protocol, the content tells you nothing. What remains is metadata — destination, port, interval, size.
Ingress Tool Transfer (T1105) is the largest single block of rules in the tactic, and the best-evidenced one, because the tool arrives over the network in the clear often enough to be seen: a fetch by curl or wget or Invoke-WebRequest from an address the machine has never used, or a binary written to disk straight after it. The download is the moment the intrusion becomes an implant rather than a mistake.
Remote Access Tools (T1219) and its remote desktop sub-technique (T1219.002) are the commercial case: TeamViewer, AnyDesk, ScreenConnect and their relatives, which most estates already own and have approved. The detection problem is not the software — it is the first run, the outbound relay connection, the account, the hour.
Fallback Channels (T1008) is what the intruder uses when the channel is closed, and Dynamic Resolution (T1568), including domain generation algorithms (T1568.002), is how a blocked destination is replaced by a fresh one every day.
Where the telemetry comes from
- Firewall and network logs (
firewall_syslog). Thesrc_iptodst_ippair, thedst_port, the timestamps: the source of the rhythm. - DNS logs. The queries and their answers. Often the only place the domain survives, because a shared hosting address says nothing about who is behind it.
- Endpoint network events and process creation (
windows_sysmon,linux_auditd). The process that opened the connection, itscommand_line, the account it ran as, and the event that ties the process to the destination.
How a rule actually reasons over this
A rule against this tactic is almost never a content rule — by the time the platform sees the traffic, the payload may be encrypted. It is a context rule: this process, on this host, to this destination, on this port, for this account. The rarity anchors that hold:
| Signal | Columns | Why it is rare |
|---|---|---|
| Process connecting out that never connects out | process_name, host_id, dst_ip | Most processes are silent or talk to a fixed set |
| Destination port never seen for that host | dst_port, host_id | The estate's port map is stable |
| DNS answer with a domain the host never asked for | src_ip, message | A workstation's resolver has a small vocabulary |
| Encoded DNS query — long or high-entropy label | message, src_ip | Ordinary lookups are short and readable |
| Same interval, same size, repeated | event_ts, dst_ip, dst_port | A beacon repeats; browsing does not |
| Remote access tool first seen on a server | process_name, host_id | Servers have no operator asking for a session |
| Download written straight after a fetch | process_name, file_path | The fetch and the drop are adjacent |
A rule for the tool layer looks like this:
title: Off-host download by a process that does not normally reach the network
status: experimental
logsource:
category: network_connection
product: windows
detection:
fetchers: 'curl.exe|wget.exe|bitsadmin.exe|Invoke-WebRequest'
downloaders: 'certutil.exe|bitsadmin.exe|mshta.exe'
selection:
process_name|contains|:
destination_port|exists: true
filter_updates:
process_name|contains: 'svchost'
condition: fetchers or downloaders
fields: [process_name, parent_process, user_name, host_id, src_ip, dst_ip, dst_port]
The important part is not the fetchers — it is that the rule is anchored on the pair of a download and a write, because T1105 is the moment the implant lands and the moment with the clearest telemetry.
The rhythm, and why a rule is not enough
Beaconing is a property of a series, not of an event, and that is where the corpus runs out. The rules catch the tool, the port and the known infrastructure; none of them can say "this connection happened 43 times, once every 60 seconds, with 4% jitter, and nothing else in the estate does that". That is a correlation over time, and it is why a programme with only signature rules misses a careful intruder using a real browser, a real certificate and a real domain.
Two consequences follow. The interval has to survive a sleep cycle — an intruder who sleeps for hours has a jitter of hours, and a threshold tuned to a 60-second beacon will not fire on a 4-hour one. And the baseline must be per host and per account, because a beaconing pattern is only remarkable against that host's own history.
Where analysts get this wrong
The first mistake is trusting the destination address. A modern command channel is hosted on shared infrastructure, so the address the machine contacts may belong to a hosting provider with thousands of tenants behind it. A rule built on a block list of addresses is a rule that decays. The domain, the certificate and the pattern are the better anchors — and the domain is exactly what the address throws away, which is why DNS logs matter here.
The second mistake is trusting the port. Traffic on 443 and 53 passes everything, and the sub-techniques that exist precisely to make it pass — tunnelling, proxies, dead drops — are the ones that will be missed. The third is forgetting the first download: T1105 happens before any of the hard-to-see behaviour begins.
Detection engineering notes
- Anchor on the pair, not the property. "Outbound connection" is not an event worth an alert. "Outbound connection from a process that has never made one, to a destination this host has never used" is.
- Keep DNS. If the only network telemetry you keep is the connection, you have thrown away the domain — and the domain identifies the adversary.
- Model the interval per host. A global baseline makes every regular client look like a beacon, and every careful beacon look like a client.
- Watch the download, not the implant. The fetch is visible; the implant is designed not to be.
- Know the legitimate remote tools.
T1219.002is not a rare technique — it is a popular one, and the estate's own approved remote support software is the baseline it must be compared against.
Where DetectMonkey fits
There is one category of telemetry an organization should think twice about before sending anywhere else: the channel the intruder uses to give instructions. It shows what the attacker wants, when they want it, and what is leaving the building in reply.
DetectMonkey processes and stores that telemetry on private servers operated by Codein in New Brunswick, Canada — the command-and-control records, the DNS queries, the process-to-network correlations — and never in a public cloud. The collectors read the endpoint and firewall streams the channel leaves behind, and the pinned corpus of 3164 Sigma rules evaluates them as the events arrive, so a new destination or a new tool is seen while the session is still open. The last 72 hours stay queryable, which is what it takes to reconstruct a channel's rhythm after the fact rather than during it.
The platform detects and alerts. It does not isolate, quarantine or block anything on your systems, and it does not change your configuration. Analysts work in English and French, and every alert carries its evidence in both. 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.