ATT&CK guide
Cloud and identity: the attack surface that is not a laptop
The shortest path into a hundred-person company is a browser tab. This is what cloud identity telemetry actually shows, why an audit trail nobody watches is not a detection, and what you can do this week.
For the owner
If your company runs on a mail suite, a shared drive and a payroll system in someone else's cloud, then your real perimeter is not the office network. It is the login box.
An intruder does not need to break into your building. They need one credential, one forgotten account, one application a staff member approved without reading, or one supplier who shares a cloud account with you. From there, the same company files, the same customer list and the same mailbox rules are reachable from anywhere on earth — and the record of what was taken sits in a log nobody opens.
The pattern that keeps repeating is not dramatic. A sign-in from an unexpected country. A new administrator added to a directory. A forwarding rule on a mailbox that had none. An application granted access to everything. Each is one line in a log, read once a year, if ever.
Three questions worth asking your own team:
- If an account of yours were used from another continent tomorrow, who would find out, and how fast?
- Who can grant an application access to your mail, and who can add an administrator to the directory?
- Do you have a record of what happened in your cloud accounts after an employee leaves, or a laptop is stolen?
The difference is not the size of the log but whether anything is watching it — and most small organisations have the logs without the watching.
For the analyst
The perimeter moved, and it moved quietly
For a hundred-seat organisation, most interesting assets are now identity objects rather than machines. The mailbox, the shared drive, the payroll spreadsheet, the code repository, the customer record: they live behind an identity provider — a service built to accept requests from the entire internet and let legitimate people in from anywhere.
That design is the problem for detection. Cloud identity is built for availability, not for noticing: a successful sign-in is a success message. The interesting fact is not "the login worked" but everything around it — the address, the device, the hour, the country, and whether this account has ever done this before.
The corpus carries meaningful coverage here, and the most-cited techniques are the ones this post is about:
| Technique | Name | Rules in the corpus |
|---|---|---|
T1078 | Valid Accounts | 51 |
T1078.004 | Valid Accounts: Cloud Accounts | 37 |
T1098 | Account Manipulation | 32 |
T1110 | Brute Force | 23 |
The telemetry that actually exists
The platform's collectors read identity audit trails directly: Microsoft 365 and Entra ID (m365_audit), Google Workspace (gworkspace), and Keycloak for organisations that run their own identity. The normalised columns are the familiar ones — user_name, src_ip, dst_ip, event_type, service_name, severity, message, event_ts — so an identity event and an endpoint event can be correlated in the same query.
The event types worth naming:
- Sign-in. Time, source address, user agent, device identifier, result, and the conditional-access decision — the raw material for the detection that matters most.
- New device, new location. Usually carried on the sign-in record itself, which is why the collector keeps those fields rather than flattening the record into a message.
- Token grant and application consent. An application asking for a permission set and a user approving it — the modern equivalent of a saved credential on a shared drive, and much harder to see.
- Privilege change. A role granted, an administrative user added, an application's access level modified, domain-wide API access granted, a service account created.
- Mailbox forwarding rule. Created, changed, or pointed at an address outside the organisation.
- Credential and second-factor changes. A second factor disabled, a recovery method changed, a password reset, a session token issued.
- Administrative actions. The console activity of an administrator: the highest-signal stream in the tenant, and the one most often left uncollected.
How a rule reasons over identity telemetry
Identity detections are, almost without exception, baseline detections. The event is unremarkable; what makes it notable is the comparison against what that account has done before.
Impossible travel (T1078.004). Two sign-ins, two source addresses with geolocations, and an elapsed time shorter than the distance between them allows. The corpus carries both a Microsoft 365 impossible-travel rule and an Azure Identity Protection equivalent. The weakness is worth stating: it depends on IP geolocation, it is defeated by an exit node in the right country, and an intruder signing in from the victim's own country never trips it.
Valid account use (T1078). The success event is identical to a legitimate one, so the rule reasons about the neighbourhood: a first sign-in from a new address, on a new device, with a new user agent, at an hour this account has never been active. Each is weak alone; together with the other telemetry on the same user_name they are strong.
Account manipulation (T1098). An allowlist problem. Which privilege changes are routine — a service account role updated by a scheduled job, a new member of a departmental group — and which are a person quietly promoting themselves? The corpus contains rules for users added to administrative roles, applications granted broader access, and domain-wide API access granted, because those precede the intruder becoming harder to remove.
Brute force (T1110). Repeated failures across accounts — a password spray pattern rather than a single-account one — and successful authentication immediately after a burst of failures for the same user.
Exfiltration through the identity. The most valuable end-to-end correlation, and one the cloud audit trail makes possible: a mail rule created on a mailbox that never had one, followed by an unusual volume of file access or an upload to an application that is not on the sanctioned list. The rule is the intent; the volume is the act.
An audit trail that exists is not a detection
The distinction the industry keeps sliding over. Having logs is a capability; having a detection is a product decision. Between the two sit three requirements, none of them free:
- Collection. The logs must be gathered continuously and delivered somewhere that is not the tenant itself. If the intruder can reach the tenant's logs, the intruder can edit them.
- Retention. Identity logs are the shortest-lived evidence in most environments — often a matter of weeks — while the question "what did this account do on the night of the incident" is asked months later.
- Attention. Somebody has to read them. A log exported to a bucket nobody queries produces the same outcome as no log at all, and the same false confidence.
An organisation can satisfy all three without buying anything: ship the identity audit trail to a system that evaluates rules as the events arrive, keep it long enough to answer last year's question, and have a named person responsible for the alerts.
What a small organisation can do in a week
- Turn on the identity logs. Microsoft 365 and Google Workspace both emit sign-in, administrative and token events; the audit trail is a configuration switch, not a purchase. Have them delivered somewhere that is not the tenant.
- Alert on four things. A new administrative role granted, an application granted access, a mail forwarding rule created, a second factor disabled. These four are quiet, rare, and cover the great majority of cloud identity intrusions.
- Remove standing administrator rights. Use time-bound elevation for the people who need it. A permanently privileged account is a credential waiting to be stolen.
- Separate the backup identity from the human identity. If the account that reaches the backups is the one in a browser tab, one stolen session reaches everything.
- Revoke what nobody uses. Dormant accounts, stale application consents and departed employees' tokens are the cheapest wins on the list, and the paths an intruder prefers.
Where analysts get it wrong
The first error is believing multi-factor authentication closes the question. It closes the password-phishing question. It does not close a stolen session, a consented application, a device already trusted, or an API token — none of which produce a sign-in event that looks wrong.
The second is treating impossible travel as the primary detection. It is a good detection and a poor perimeter: it misses the attacker who signs in from the victim's own country, and it is trivial to defeat with an exit node in the right place.
The third is treating "no unusual application" as proof. The attacker who wants a durable, quiet path into a mailbox will consent to an application that looks ordinary, or use one that is already sanctioned. The signal is the combination — a consent granted, from an unusual account, granting scopes it does not need — not the vendor name.
The fourth is the one that hurts most in a small organisation: the identity that was compromised belonged to somebody nobody watches. A service account, a shared mailbox, a departed employee's account that was never closed. Nothing fires on an account that has no baseline, because there is no baseline to be unusual against.
Where DetectMonkey fits
Identity audit trails are among the most revealing telemetry a company collects and the most sensitive data it holds: who signed in, from where, on which device, and what they changed. That combination is why residency is a real question here rather than a marketing line.
DetectMonkey ingests the identity audit trails alongside the endpoint and perimeter telemetry, evaluates the pinned corpus of 3164 rules against them as the events arrive rather than after they are indexed, and keeps the last 72 hours queryable so that an impossible-travel alert, an administrator change and a mail rule can be read as one story rather than three tickets.
The platform detects and alerts. It does not isolate, quarantine, block or remediate anything on your systems, and it makes no change to 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, with every alert and ticket carrying both.
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.