Detection
A custom rule for low-and-slow SSH guessing
The built-in brute-force rule missed a slowed-down SSH attack. I wrote a custom Wazuh rule that catches it, and proved it fires.
Date: 22 October 2026
Type: Detection engineering (find a gap, write a rule, prove it)
Outcome: The built-in brute-force rule (5712) was evaded by a slowed-down attack. My custom rule (100100) caught it at level 10.
1. Hypothesis: the gap
Built-in rule 5712 fires on 8 invalid-user attempts from the same IP within 120 s (Case 002). An attacker who knows this can stay under it by:
- Slowing down: never 8 within any 2-minute window.
- Spreading out: a few attempts from each of many IPs (
same_source_ipignores them), as in password spraying (T1110.003).
2. The rule
detections/wazuh/local_rules.xml
<rule id="100100" level="10" frequency="12" timeframe="300" ignore="120">
<if_matched_sid>5710</if_matched_sid>
<description>sshd: low-and-slow or distributed SSH guessing - 12+ invalid-user attempts in 5 minutes (any source)</description>
<mitre><id>T1110</id></mitre>
</rule>| Design choice | Why |
|---|---|
| A 5-minute window (vs 2) | Catches attackers who pace their attempts |
No same_source_ip |
Catches attempts spread across many IPs |
| Threshold 12 (vs 8) | The longer window needs a higher bar, so normal mistakes don't trigger it |
ignore="120" |
One attack produces one alert, not a flood |
| ID 100100 | The custom range (100000+) never clashes with Wazuh's built-in rules |
3. True-positive test: a low-and-slow attack
# 7 attempts, pause just over 5712's 120 s window, 7 more
for i in $(seq 1 7); do ssh -o BatchMode=yes -o ConnectTimeout=5 fakeuser@10.0.2.10; done; \
echo "Pausing 130 seconds..."; sleep 130; \
for i in $(seq 1 7); do ssh -o BatchMode=yes -o ConnectTimeout=5 fakeuser@10.0.2.10; done| Search (last 15 min) | Result |
|---|---|
rule.id:5710 |
13 alerts in two bursts (16:44:29 and 16:46:41) |
rule.id:5712 (built-in) |
No results: evaded |
rule.id:100100 (custom) |
1 alert, level 10 at 16:46:41, data.srcip 10.0.1.101, previous_output listing the attempts |
Note: 14 attempts were sent but 13 were logged. Open question: one connection probably never reached the login stage. To check, count the Permission denied lines on Kali.
4. False-positive test: normal mistakes
Three invalid-user attempts (a user mistyping), sent more than 5 minutes after the attack. Expected and observed: 3 new 5710 alerts, and no new 100100 or 5712 alert. (To be confirmed with a screenshot.)
5. Limitations and next steps
- Lab limitation: both bursts came from one IP, so this proves the slow evasion. The distributed case (many IPs) still needs testing from a second attacking machine.
- Tuning: in production, a busy server might see 12 honest failures in 5 minutes. Baseline normal failure rates before deploying, then tune the threshold or limit it to sensitive servers.
- Next rule ideas: many different usernames from one source (username enumeration), and a successful login shortly after many failures (likely compromise).
Lessons
- Detection engineering is a loop: hypothesis → rule → true-positive test → false-positive test → document.
- Reading a rule's conditions reveals how attackers evade it.
- Every threshold trades missed attacks against false alarms. State the trade-off.
Evidence
| Screenshot | Shows |
|---|---|
| Rule loaded | 100100 in local_rules.xml, level 10 |
| Two bursts | 5710 alerts at 16:44 and 16:46 |
| 5712 evaded | No built-in brute-force alert |
| 100100 fired | The custom rule at level 10 |
| Alert detail | Source IP, frequency 12, previous_output |