A finished detection record
A detection record has a section for each thing a reviewer or responder needs: goal, categorization, strategy abstract, technical context, blind spots and assumptions, false positives, validation, priority, response, additional resources, and the retirement record. The two records below, one Windows and one cloud identity, show the shape in more than one domain. The format follows Palantir's Alerting and Detection Strategy.
Remember
- Write four sections at hypothesis: goal, categorization, strategy abstract, and blind spots. You fill in the rest during research and build.
- In the blind spots section, name a specific case the detection misses.
- Write response steps that someone who has never seen the detection could follow at 3 a.m.
- Date each assumption about the environment, so a later reader knows what normal meant at the time.
- Put each exclusion in an allowlist with a ticket id, an owner, and an expiry date.
On this page
DET-0001 Kerberoasting: RC4 service ticket request for a user account
Goal
What an alert from this means, in one sentence a responder can act on.Catch an attacker collecting Kerberos service tickets for services that run under user accounts, so that the tickets can be cracked offline to recover the account passwords (ATT&CK T1558.003).
Categorization
The ATT&CK technique, so coverage can be counted and duplicates found.Credential Access. T1558.003, Steal or Forge Kerberos Tickets: Kerberoasting.
Strategy Abstract
The logic in prose, so a reviewer can disagree without reading the query.Hypothesis: when an attacker runs Kerberoasting, the domain controller writes event 4769 (a service ticket was requested). The event carries encryption type 0x17 (RC4-HMAC) and status 0x0 (success), for a service run by a person or a service account. The detection alerts on that combination and ignores computer accounts (names ending in $) and the krbtgt account. Log source: the Windows Security log on every domain controller, event 4769.
Technical Context
Why the behavior looks like this, and what normal traffic in it looks like.Any domain user can ask for a service ticket for any service in the domain. The ticket is encrypted with the service account's password hash, so if the ticket was issued with RC4, the attacker can guess passwords against it offline at high speed. Attackers ask for RC4 on purpose: AES tickets crack far more slowly. The 7-day count during research shows how many legitimate RC4 requests the environment still produces. Older applications, and accounts that never had "This account supports Kerberos AES" set, are the usual sources.
Plan approved by: nobody; no lab was provisioned. The fixtures come from a public recording, OTRF Security-Datasets SDWIN-201022042947, in which PurpleSharp ran T1558.003 against a lab domain on 2020-10-22. Run tag: SDWIN-201022042947. The recording's 17 event 4769 records are the run; 5 of them are RC4 requests for user-run services. Baseline: not measured. The kit's example has no production environment to count. A team copying this detection counts 4769 events with encryption type 0x17 over 7 days of its own domain controllers before setting the threshold. Teardown verified: not applicable; nothing was provisioned.
Blind Spots and Assumptions
What this will miss. The section reviewers argue with.- Every domain controller has the "Audit Kerberos Service Ticket Operations" policy on, and its events reach the SIEM.
- An attacker who accepts AES tickets and cracks them slowly is not seen by this detection.
- A request handled by a domain controller whose logs are missing is not seen. DET-0000, the silence alert, is what catches the missing logs.
False Positives
What fires it that is not an attack, and how the exclusion is made.Older applications that still ask for RC4 tickets. After the 7-day count, exclude them by ServiceName in a pull request, never in the SIEM's interface.
Validation
How somebody proves it fires, and where the fixtures live.The fixtures in tests/events.jsonl are five real event 4769 records from the SDWIN-201022042947 recording: two RC4 requests for user-run services (sysmonsvc, nxlogsvc) the detection must catch, and three it must ignore (two machine accounts and krbtgt). The recording has no RC4 request for a machine account, so the machine-account filter is exercised only together with the AES filter. A request with both RC4 and a machine account would still be ignored, but no fixture proves it. CI replays them on the SQLite engine on every change and diffs every converted query against its approved copy. On 2026-10-15 the Splunk container and Kusto emulator adapters ran them end to end by hand and returned the same two events. This team deploys to Splunk and Sentinel, so those two jobs become required checks once they run in CI.
Priority
The score and where the numbers behind it are recorded.High. Scores 2.40 on the balanced preset; score.yml carries the six inputs and the note behind each one.
Response
The ordered steps to take at three in the morning.- Find the requesting account (TargetUserName) and where the request came from (IpAddress).
- List every ServiceName that account requested in the same hour.
- Check the requesting account and machine against the allowlist of legacy applications before anything is reset.
- Look at the requesting machine for the tool that made the request.
- Change the passwords of the service accounts behind those services; move them to group managed service accounts where possible.
Additional Resources
Where the claims above came from.- ATT&CK T1558.003
- Microsoft audit policy: Audit Kerberos Service Ticket Operations
Retirement record
Empty until it retires, then the trigger that fired and what was measured.DET-0001 is at the build stage and has no retirement to record.
DET-0101 Password spray against Entra ID sign-in
Goal
What an alert from this means, in one sentence a responder can act on.Catch an attacker trying a small number of common passwords against many Entra ID accounts from one source, so the attempt is stopped before a single account is taken (ATT&CK T1110.003).
Categorization
The ATT&CK technique, so coverage can be counted and duplicates found.Credential Access. T1110.003, Brute Force: Password Spraying.
Strategy Abstract
The logic in prose, so a reviewer can disagree without reading the query.Hypothesis: when an attacker sprays passwords at the tenant, the Entra ID sign-in log records many distinct principals failing from one source inside a short window. Sign-in logs record one row per authentication attempt with a result code. Error 50126 is a wrong password for a real account. A password spray looks like one actor, many distinct accounts, few attempts each, over a short window. The rule counts distinct failed principals per source in a rolling hour and alerts above the threshold, then reports whether any account in the same set succeeded afterwards. A spray followed by a success is an incident. Log source: the Entra ID sign-in log, streamed through diagnostic settings.
Technical Context
Why the behavior looks like this, and what normal traffic in it looks like.Normal sign-in failure in a corporate tenant is one account failing repeatedly from its own device after a password change or a stale token on a phone. That is the opposite shape: few principals, many attempts. Spraying inverts it deliberately to stay under per-account lockout. That is why the rule counts distinct principals.
Two sources of legitimate inversion exist. A misconfigured application using a service principal can fail against many mailboxes at once, and a security scanner running credential checks will do it on purpose. Both are identifiable by their source and belong in the allowlist with an owner and an expiry.
Sign-in logs arrive through the diagnostic settings pipeline with a delay of under 5 minutes at the 95th percentile, measured over the 7-day baseline. That delay sets the floor under this detection's latency, whatever the schedule is.
Blind Spots and Assumptions
What this will miss. The section reviewers argue with.It assumes the attacker reuses one source address or one autonomous system. A spray distributed across residential proxies, one attempt per address, will not cross the threshold and this detection will not see it. The distributed case needs an identity-side signal instead: failures per principal across sources. That is a separate detection, on the backlog.
Conditional access does not hide the failures: Entra ID checks the password before any policy runs, so a wrong guess logs 50126 whatever the policy says. A correct guess blocked by policy logs 53003 instead of a success, so the follow-up success check treats 53003 from the spraying source as a hit.
It sees only Entra ID. Spraying against on-premises AD, ADFS, or a third party federated to the tenant produces different logs.
False Positives
What fires it that is not an attack, and how the exclusion is made.A service principal authenticating on behalf of many users after a credential rotation. A vulnerability scanner performing credential checks. A load test against a non-production tenant that shares the log stream. All three are named exclusions in the allowlist table, each with a ticket id, an owner, and an expiry date.
Validation
How somebody proves it fires, and where the fixtures live.True-positive fixture: 14 sign-in events, 12 distinct principals, one source address, result 50126, inside 8 minutes, captured from an authorized credential-access exercise on 2026-08-14.
False-positive fixture: 8 sign-in events, 1 principal, one source, result 50126, spread over an hour. That is a person with a stale password on a phone, and it stays under the tenant's smart lockout threshold of 10, after which the result code changes to 50053.
Both are written for a replay with the threshold at its default and at both bounds. The kit doesn't replay at the bounds yet, and the rule file and fixtures aren't in the reference repository, so neither has run.
Priority
The score and where the numbers behind it are recorded.Score 2.60 on the balanced preset. threat_relevance 5, impact 4, robustness 4, telemetry_readiness 5, build_effort 2, run_cost 3. Marginal coverage: first, no live detection covers T1110.003. The six evidence lines are on the score record.
Response
The ordered steps to take at three in the morning.- Confirm the source: is it a known service principal, a scanner, or an unrecognized address.
- List the principals in the set and check each for a successful sign-in from the same source in the following hour.
- If any succeeded, this is an incident. Revoke the sessions and reset that account before anything else.
- If none succeeded, block or throttle the source, confirm the affected accounts are covered by MFA, and check whether any of them are exempt from conditional access.
- Record the count of principals, the source, and whether any success followed, because those three numbers decide whether the threshold is right.
Additional Resources
Where the claims above came from.Entra ID sign-in error codes reference for 50126. The 7-day baseline query and its output, stored with the detection. The authorized exercise ticket from 2026-08-14 that produced the true-positive fixture.
Retirement record
Empty until it retires, then the trigger that fired and what was measured.Empty. If it retires, this section records which trigger fired, what was measured, and the link to whatever replaced it.
The blank version
Copy this into a file next to the detection. A new detection writes goal, categorization, strategy abstract, and blind spots at hypothesis, and the rest as it moves through the stages. From build on, lint fails a record missing any heading. From validate on, it also fails any empty section but the retirement record. A detection already live whose record sits on a wiki or in its own description field can start with four sections: goal, categorization, blind spots and assumptions, and response.
# DET-XXXX <name>
## Goal
## Categorization
## Strategy Abstract
## Technical Context
## Blind Spots and Assumptions
## False Positives
## Validation
## Priority
## Response
## Additional Resources
## Retirement record