Canadian data residency · made, hosted, and operated in Canada

Early access — 90 days →

MITRE ATT&CK Enterprise · TA0005

Stealth: how an intruder makes the work look normal

StealthTA0005

The quiet phase of an intrusion is the expensive one. This is the tactic that describes it — what it costs, what it looks like from the outside, and the telemetry that survives it.

Codein Cybersecurity November 22, 2026 10 min read

For the owner

An intruder who wants time does not make a mess. They make the work look like the work your staff does every day: a tool running on a schedule, a script started by another script, a file copied in by an update, a log cleared because somebody was tidying up.

That is the quiet phase of an intrusion, and it is the most expensive phase. Nothing looks wrong, so nobody asks. Meanwhile the intruder is inside the systems that hold your customer records, your payroll, your operational technology, and the credentials that open all of it. The damage is not done during the break-in; it is done during the weeks when nobody is looking.

The visible effect, from the outside, is a company that looks exactly like it did last month — until the day it does not, and by then the story is hard to reconstruct because the evidence was tidied away as part of the intrusion.

Three questions are worth asking your own team:

  1. Do we know what normal looks like on our systems? A scheduled job, a script, a service, a file name — do we know which ones are ours and which ones appeared last week? If the answer is no, nothing can be flagged as out of place, because there is no baseline to compare it to.
  2. When a log or a record is deleted or cleared, who finds out? Deleting history is unusual, and unusual is cheap to detect. Most organisations never find out at all.
  3. If our logging were partly broken tomorrow, would we know before the intrusion did? A team that cannot answer this is running a system where the quiet phase is invisible by default.

None of this needs new spending. It needs the decision to treat the boring, ordinary activity on your systems as worth recording, and to alert on the things that should not happen. That is where the quiet phase ends.

For the analyst

The tactic

MITRE ATT&CK files this tactic as TA0005, named Stealth. It was filed for years under the name Defense Evasion, and you will still find that name in older material and in vendor taxonomies; the substance has not changed. The adversary's objective here is time. Everything done under this tactic buys minutes or hours of not being noticed, so the actions are, by design, the ones that look like ordinary administration.

The matrix files 151 techniques under it, and the pinned SigmaHQ corpus carries 848 rules that name techniques in this tactic. It is one of the two largest tactic blocks in the corpus, and the reason is structural: hiding is not one technique, it is a menu of them.

The techniques the corpus detects

TechniqueNameRules in the corpus
T1218System Binary Proxy Execution140
T1027Obfuscated Files or Information88
T1574.001Hijack Execution Flow: DLL80
T1078Valid Accounts51
T1202Indirect Command Execution40
T1036Masquerading38
T1078.004Valid Accounts: Cloud Accounts37
T1218.011System Binary Proxy Execution: Rundll3227
T1036.003Masquerading: Rename Legitimate Utilities26
T1055Process Injection26
T1564.004Hide Artifacts: NTFS File Attributes22
T1070Indicator Removal20
T1127Trusted Developer Utilities Proxy Execution19
T1036.005Masquerading: Match Legitimate Resource Name or Location17
T1218.010System Binary Proxy Execution: Regsvr3217
T1197BITS Jobs15
T1140Deobfuscate/Decode Files or Information14
T1070.004Indicator Removal: File Deletion13
T1216System Script Proxy Execution13
T1574.011Hijack Execution Flow: Services Registry Permissions Weakness11
T1564Hide Artifacts10
T1027.010Obfuscated Files or Information: Command Obfuscation9
T1070.003Indicator Removal: Clear Command History9
T1134.001Access Token Manipulation: Token Impersonation/Theft9
T1218.007System Binary Proxy Execution: Msiexec9
T1218.008System Binary Proxy Execution: Odbcconf8
T1564.003Hide Artifacts: Hidden Window8
T1218.003System Binary Proxy Execution: CMSTP7
T1218.005System Binary Proxy Execution: Mshta7
T1564.001Hide Artifacts: Hidden Files and Directories7
T1055.001Process Injection: Dynamic-link Library Injection6
T1134.002Access Token Manipulation: Create Process with Token6
T1027.003Obfuscated Files or Information: Steganography5
T1027.004Obfuscated Files or Information: Compile After Delivery5
T1070.006Indicator Removal: Timestomp5
T1078.002Valid Accounts: Domain Accounts5
T1078.003Valid Accounts: Local Accounts5
T1218.001System Binary Proxy Execution: Compiled HTML File5
T1220XSL Script Processing5
T1574Hijack Execution Flow5
T1027.005Obfuscated Files or Information: Indicator Removal from Tools4
T1055.012Process Injection: Process Hollowing4
T1070.005Indicator Removal: Network Share Connection Removal4
T1134Access Token Manipulation4
T1134.003Access Token Manipulation: Make and Impersonate Token4
T1211Exploitation for Stealth4
T1218.009System Binary Proxy Execution: Regsvcs/Regasm4
T1564.002Hide Artifacts: Hidden Users4
T1027.001Obfuscated Files or Information: Binary Padding3
T1036.002Masquerading: Right-to-Left Override3
T1036.007Masquerading: Double File Extension3
T1078.001Valid Accounts: Default Accounts3
T1497.001Virtualization/Sandbox Evasion: System Checks3
T1620Reflective Code Loading3
T1055.003Process Injection: Thread Execution Hijacking2
T1055.009Process Injection: Proc Memory2
T1216.001System Script Proxy Execution: PubPrn2
T1218.013System Binary Proxy Execution: Mavinject2
T1218.014System Binary Proxy Execution: MMC2
T1542.001Pre-OS Boot: System Firmware2
T1564.006Hide Artifacts: Run Virtual Instance2
T1574.005Hijack Execution Flow: Executable Installer File Permissions Weakness2
T1574.006Hijack Execution Flow: Dynamic Linker Hijacking2
T1574.007Hijack Execution Flow: Path Interception by PATH Environment Variable2
T1574.012Hijack Execution Flow: COR_PROFILER2
T1006Direct Volume Access1
T1014Rootkit1
T1027.002Obfuscated Files or Information: Software Packing1
T1027.009Obfuscated Files or Information: Embedded Payloads1
T1027.011Obfuscated Files or Information: Fileless Storage1
T1036.004Masquerading: Masquerade Task or Service1
T1036.006Masquerading: Space after Filename1
T1055.011Process Injection: Extra Window Memory Injection1
T1127.001Trusted Developer Utilities Proxy Execution: MSBuild1
T1134.004Access Token Manipulation: Parent PID Spoofing1
T1134.005Access Token Manipulation: SID-History Injection1
T1218.002System Binary Proxy Execution: Control Panel1
T1221Template Injection1
T1542.003Pre-OS Boot: Bootkit1
T1574.008Hijack Execution Flow: Path Interception by Search Order Hijacking1
T1622Debugger Evasion1

Five families of hiding

Proxy execution (T1218 and its sub-techniques). The attacker runs a signed Microsoft binary — rundll32 (T1218.011), regsvr32 (T1218.010), mshta (T1218.005), msiexec (T1218.007), cmstp (T1218.003), regsvcs (T1218.009), odbcconf — and passes their payload as an argument. The point is that the process tree now contains a legitimate, signed executable, so anything that alerts on "unsigned binary ran" sees nothing. The signature is therefore not in the binary. It is in the argument: rundll32 with a URL or a javascript: string, regsvr32 /i: pointing at a scriptlet in a temporary directory, mshta with a remote document, cmstp with a configuration file from the internet. The normalised fields that carry it are process_name, command_line, parent_process and file_path.

Obfuscation and masquerading (T1027, T1027.010, T1027.003, T1027.004, T1027.011, T1140, T1036, T1036.002, T1036.003, T1036.005, T1036.007). Encoded commands, packed or padded binaries, payloads embedded in images, code compiled only after it arrives, command lines assembled at runtime. Masquerading is the same idea applied to names: rename a real utility (T1036.003), copy it into a directory where it should not be (T1036.005), put a reversed character in the name so it reads differently (T1036.002), append a second extension (T1036.007). The useful asymmetry is that obfuscation leaves a footprint on the way in and on the way out: something decoded at runtime (T1140) usually means a payload was written and then executed — a file-creation event followed by a process-creation event.

Hijacking the loader (T1574, T1574.001, T1574.006, T1574.007, T1574.011). The attacker does not run their own program; they arrange for a legitimate program to load theirs, by leaving a library in a directory searched first, by weakening the permissions on a service's directory, or by replacing a dynamic linker. Detection here is not about processes at all — it is about image loads: a signed system process loading a library from a user-writable path. In the normalised columns that is process_name, file_path and event_type for image load.

Removing and hiding the evidence (T1070, T1070.003, T1070.004, T1070.005, T1070.006, T1564, T1564.001, T1564.003, T1564.004). Deleting the files they dropped, clearing command history, removing the network share connection that shows where they came from, resetting timestamps, marking a file hidden, running a console window without a window. These are the highest-value detections in the tactic and the cheapest to build, because the operations are rare in normal operation. A command history being cleared is not something that happens by accident.

Using legitimate access and legitimate tools (T1078, T1078.004, T1127, T1216, T1202, T1197, T1134, T1134.001, T1134.002). Valid Accounts, trusted developer utilities such as MSBuild, system script proxy execution, indirect command execution, background intelligent transfer jobs for a download that never looks like a download, and token theft to run as somebody else. The common thread is that a component which is allowed to perform the action performs it, so the question for the rule is not "is this allowed" but "is this what this account does on this host".

Where the telemetry comes from

  • windows_sysmon — process creation with the full command line, image load, file deletion, registry value sets. This is the primary source for the proxy-execution, masquerading and indicator-removal families. Fields: process_name, parent_process, command_line, file_path, event_type.
  • linux_auditd — execution records with the argument vector, unlink and rename operations, privilege transitions. The command_line field carries the argument vector, where the encoded and unusual arguments live.
  • firewall_syslog — the connection the download or the payload fetch made. When the endpoint record is missing or was removed, the network record is usually what is left.
  • keycloak, m365_audit, gworkspace, cloud — the identity side. A user whose activity suddenly matches a pattern they have never matched before is visible even when the endpoint record was cleaned away.
  • source_type and severity matter for triage: the same rule firing on an endpoint event and on an identity event are different alerts.

What the numbers do and do not tell you

848 rules is a large number, and it should not be read as "we cover all of this". Coverage inside the tactic is uneven, and not because of rule quality. The families whose artefacts are visible in ordinary endpoint telemetry — proxy execution, masquerading, indicator removal — carry most of the rules. The families where the attacker removes the very record you would detect them from carry one or two rules each, because there is little left to write a rule against. The honest reading: strong coverage of the ways an intruder tries to hide in plain sight, thin coverage of the ways they remove the evidence entirely. The second category is not solved by more rules; it is solved by having a second source of truth the attacker did not know about.

Detection engineering notes

  • Anchor on the argument, not the binary. A rule that matches process_name=rundll32 is noise. A rule that matches rundll32 with an argument that is a URL, or a scriptlet, or an export from a user-writable directory, is a detection.
  • Deletion is rarer than creation. Alerting on the removal of command history, of recent temporary executables, or of event logs is close to free in false-positive terms and catches a disproportionate share of intrusions.
  • Image loads deserve more attention than they get. T1574 is invisible to process-creation logging entirely, and the field that finds it — file_path pointing into a writable directory — is a normalised column every event already carries.
  • Watch the gaps. Absence is signal: a host that normally produces process creation events and produces none for an hour is telling you something the individual events cannot.
  • Tune against your own administration, not against theory. Vendor updaters, installers and backup agents legitimately use rundll32, msiexec, PowerShell and scheduled tasks. A rule set that fires on all of them teaches your team to ignore the tactic entirely.
  • Correlate across hosts. The same masqueraded utility name appearing on one host is a curiosity; the same name on four hosts with the same unusual parent process is a campaign.

Where analysts get this wrong, and what a false negative looks like

The most common error is treating stealth as the objective. It is not the objective; it is the purchase. The intruder hides in order to do something later, so an alert that only says "this process was masquerading" leaves the analyst with a technique and no story. The second common error is assuming that a good rule set means the intrusion is visible. A rule that reads process creation cannot see an intruder who removed process creation.

That is the false negative to internalise: the intruder does not merely delete files after the fact — they also disable the recording, adjust the settings so the events do not flow, and act inside a window where the endpoint is quiet. Then the endpoint record does not exist. What remains is the identity stream, the network records and the behaviour of the account itself, which is why the coverage question for this tactic is never answered by the endpoint alone.

Where DetectMonkey fits

Stealth is where the analytical value of a detection platform is decided. When the intruder is deliberately working to look ordinary, the question is no longer "what did this machine do" but "what does this account do that it has never done before" — and that question is answered by correlation across hosts, across sources and across time, not by any single rule. The corpus of 3164 pinned detection rules covers the common hiding techniques; the 72-hour window behind every alert keeps the surrounding context available, so an analyst can ask what else that account touched in the same hour rather than what one event said.

DetectMonkey is that layer. The collectors carry endpoint, identity and firewall telemetry into one normalised event stream, the rules are evaluated as the events arrive, and every alert carries its evidence in English and French. The platform detects and alerts. It does not isolate, quarantine, block or remediate anything on your systems, and it does not change your configuration. Telemetry is processed and stored on private servers operated by Codein in New Brunswick, Canada — not in a public cloud.

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.