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

Early access — 90 days →

MITRE ATT&CK Enterprise · TA0043

Reconnaissance: how an attacker studies you first

ReconnaissanceTA0043

The most expensive part of an attack is the part where the attacker is still learning about you. This is what that research looks like, what it costs, and the telemetry that catches it.

Codein Cybersecurity November 22, 2026 6 min read

For the owner

The first thing an attacker does is not break in. It is read.

Everything they need to plan an attack against you is, to a surprising degree, already published. Your job postings name your software and your reporting lines. Your website names your products. Your staff write about their work. A supplier's case study about you names your partners. None of that is a leak, and all of it is reconnaissance.

What it looks like from the outside is rarely dramatic. It is an email that quotes an internal project name that only appears in one of your own job ads. It is an invoice that arrives from a supplier's genuine mailbox. It is a login attempt against an account whose name exists only on your public site. It is a recruiter who knows your team's headcount before you do.

The business cost lands later and lands somewhere else: a convincing letter to your finance team, an account takeover, an intruder who knows which door is weak before they ever knock on it. Research is the cheapest phase of an attack, and the only one where you are usually not watching.

Three questions are worth asking your own team, and they are worth more than any tool you might buy:

  1. What does a stranger learn about us in one hour? Have somebody who did not build your systems actually try it, and write down what they find.
  2. Which of our documents carry information about our infrastructure in their properties rather than their content? A spreadsheet, a PDF, a slide deck with an author name and a template can say more than you think.
  3. When we publish something, do we treat it as a document or as something an attacker will read? Every job ad, case study and annual report is reconnaissance material for somebody.

None of this is expensive to fix. Most of it is a decision about what to publish and a habit of looking at your own public footprint with an outsider's eyes. The reason it matters is that reconnaissance is the phase where the attacker is cheapest and you are most exposed.

For the analyst

The tactic

MITRE ATT&CK files Reconnaissance under TA0043. It is the "before" of an intrusion: the phase in which the adversary learns about the target rather than touches it. The matrix files 46 techniques under it, and the pinned SigmaHQ corpus carries 17 rules that detect techniques in this tactic.

That ratio is the first thing to understand about this tactic. Almost all of it happens somewhere you do not own — on a public website, in a code repository, in a leaked contact list. A rule cannot detect a thing that leaves no telemetry in your estate, so the rules that exist are the ones that watch the edge, where reconnaissance becomes visible.

What the corpus detects

TechniqueNameRules in the corpus
T1592.004Gather Victim Host Information: Client Configurations3
T1595Active Scanning3
T1590Gather Victim Network Information2
T1591.004Gather Victim Org Information: Identify Roles2
T1593.003Search Open Websites/Domains: Code Repositories2
T1589Gather Victim Identity Information1
T1589.002Gather Victim Identity Information: Email Addresses1
T1590.001Gather Victim Network Information: Domain Properties1
T1590.002Gather Victim Network Information: DNS1
T1595.002Active Scanning: Vulnerability Scanning1

Six families sit under this tactic, splitting by what the adversary collects and how they collect it.

Gather victim identity information (T1589, T1589.002). Names, email addresses, roles. This is the family that produces spearphishing later, and it is collected in bulk from directories, conference sites, press releases and the corporate social accounts your staff keep without being told not to.

Gather victim network information (T1590, T1590.001, T1590.002). What your public DNS says, what your domains resolve to, what mail records exist, whether there is a partner you did not think of. Both sub-techniques with rules — domain properties and DNS — are cheap to run against a domain you own.

Gather victim host information (T1592.004). Client configurations — the most rules in this table, and the most interesting. It is about the configuration of the user's machine: browser extensions, installed agents, proxy settings. An adversary collecting this is asking "what will this person's browser do when I send them a file", and the answer decides whether they send a link, a macro or a script.

Gather victim organization information (T1591.004). Identify roles: who does what, who to impersonate, which department handles invoices.

Active scanning (T1595, T1595.002). Probing your public services to find out what answers. T1595.002 (vulnerability scanning) is the version where the probe carries a payload rather than just a connection.

Search open websites and domains (T1593.003). Code repositories: public forums, paste sites, credential dumps, and — pointedly — your own source control. A repository with a hard-coded key or an internal hostname in it is reconnaissance material you published yourself.

Where the telemetry comes from

Reconnaissance reaches you through the edge, not through the endpoint. The collectors produce the following source types, and each one contributes a different part of the picture:

Source typeWhat it contributesThe reconnaissance signal
firewall_syslogPerimeter connectionssrc_ip touching many dst_port values, or many dst_ip values, in a short window
windows_sysmonProcess creation on the hostAn internal tool enumerating the estate, or a process probing network neighbours
linux_auditdProcess creation on the hostThe same, from a Linux host, including shell-outs to enumeration utilities
keycloakIdentity eventsFailed sign-ins against account names that exist only in your public material
m365_auditCloud identity eventsSign-in attempts, new device registrations, application consents
gworkspaceCloud identity eventsThe same signals from the Google workspace side
cloudProvider-level eventsObject access and egress patterns from a compromised tenant

Every one of them lands in the same normalised columns — tenant_id, host_id, event_ts, source_type, event_type, severity, src_ip, dst_ip, dst_port, user_name, process_name, parent_process, command_line, file_path — which is the point. A rule that reasons over src_ip and dst_port does not care whether the event came from a firewall or from a cloud audit trail.

What makes an event rare, and what makes it wrong

A single failed connection is common — the whole internet fails connections at your perimeter every second. That is why "a connection was refused" is not a rule and never will be.

What is rare is the shape of the event:

  • Fan-out. One src_ip, many dst_port values on one dst_ip. That is a scan.
  • Breadth. One src_ip, many dst_ip values across a range you own. That is a sweep of your address space.
  • Enumeration. One source, many distinct paths under one web server, with the responses changing from one path to the next. That is a crawl for something specific.
  • A name that should not exist. A sign-in attempt for an account name that only appears in your public job ads, arriving at an hour your staff do not work.

And what makes it wrong — the noise that ends a rule's life in an alert queue — is almost always legitimate scanning:

  • Contracted vulnerability scanners and the monitoring tools your own vendor runs.
  • Cloud provider health checks, which sweep ranges of addresses that include yours.
  • Mail servers probing for open relays.

The distinguishing feature is not the source address; it is whether the intent is followed by a second, different kind of event.

Detection engineering notes

  • Anchor on the aggregate, not the event. The unit of detection is a window: one source, a fan-out threshold, over minutes. A single-record rule for this tactic will either fire forever or never.
  • Correlate the perimeter with the identity. This is the highest-value rule shape in the tactic: a port scan against your public range from address X, followed within hours by a sign-in attempt against a named account from address X. The scan alone is trivia; the pair is an intrusion.
  • Expect the adversary to be slow. A patient attacker spreads a scan over days and rotates source addresses. Rate-based detection on a rotating set is much harder to write and much harder to evade than a blocklist, which is why a source-based rule degrades to nothing on day two.
  • Allowlist the scanners you pay for, by source, per tenant, with an expiry. A permanent allowlist for a scanner that stops being run is a hole with a maintenance ticket nobody filed.
  • Watch the identity layer for names you published. You know which of your identifiers are public. A sign-in attempt against one of them from an address with no history on the account is worth a look even if it fails.

Where analysts get this wrong

Three mistakes are common here.

The first is treating reconnaissance as unactionable because "nothing was compromised". That inverts the logic. Reconnaissance is the phase where the adversary's cost is lowest and their information gain is highest. A detected scan that led to a failed sign-in is a cheap win; the same scan that went unnoticed is the beginning of everything that follows.

The second is expecting endpoint telemetry. Most reconnaissance leaves nothing on your systems at all. It is done from rented infrastructure against public services. If your detection for this tactic lives only on the workstation, you are blind to the phase that matters.

The third is alerting on volume alone, which produces the fatigue that makes a team ignore the tactic. A big block of traffic from one address is usually a scanner. The alert worth waking someone for shows scanning followed by an attempt to use what was found.

A false negative in this tactic has a recognisable shape: an attacker enumerates your public surface slowly, one port every few hours, across three weeks, from addresses that rotate. Every individual event is a normal failed connection. No single event is rare, and the only signal that would have caught it — the rate across a set of sources against a set of ports — is exactly the kind of aggregation most rule sets never get written.

Where DetectMonkey fits

Reconnaissance is the phase with the least telemetry and the most influence on everything after it. The evidence that it happened — who was looking, what was probed, which names were tried — lives in perimeter logs and identity records, and that is some of the most sensitive telemetry an organisation produces. It is processed and stored on private servers operated by Codein in New Brunswick, Canada, not in a public cloud.

DetectMonkey detects and alerts on it. The pinned corpus of 3164 rules is evaluated as events arrive rather than after indexing, so a scan that starts at two in the morning is an alert at two in the morning and not a report the next morning. It does not isolate, quarantine or block anything on your systems, and it changes no configuration on them. The analyst work happens in English and French, and every alert carries the evidence that produced it.

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.