返回提示词目录第 1173 页

安全检测规则降噪与验证方案

根据告警样本归纳噪声来源,输出调整后的检测规则、潜在盲区说明及回放验证计划。

完整提示词

<context>
A detection that cries wolf trains analysts to close it unread, which is worse than no detection. The usual fixes make it worse in a different way: excluding a process by name (attackers rename binaries), excluding a whole host group or all admin accounts (the accounts attackers want most), raising a threshold until the true positives also disappear, or suppressing the rule entirely. Good tuning starts by grouping the noise into a few clusters with a shared cause, then excludes each cluster on attributes an attacker cannot easily copy: full path plus signer, parent process, a specific service account on a specific host, a command line pattern anchored at both ends. Every exclusion is written down with an owner and a review date, and the rule is re-tested to prove the attack it exists for still fires.
</context>

<task>
Tune this detection{{#rule_language}} ({{rule_language}}){{/rule_language}}.

<rule>
{{rule}}
</rule>

<alert_sample>
{{alert_sample}}
</alert_sample>

1. Restate what attack behaviour the rule is meant to catch. If the intent is unclear from the rule, say what you assumed.
2. Group the sample into noise clusters: the shared cause (a backup agent, a deployment tool, an IT script, a scanner), the number of alerts, and the fields that identify the cluster. Keep the true positives and anything undetermined in their own group.
3. For each cluster, list tuning options and rate each for how much noise it removes and how easily an attacker could hide inside it: narrow exclusion on stable attributes, adding a required condition to the main logic, splitting into a high-fidelity alert and a lower-severity hunting query, enrichment that auto-closes known-good, or fixing the noisy source itself (for example asking IT to sign their script).
4. Recommend one change set and write the tuned rule or the exclusion in the rule's own syntax, with comments naming each exclusion's reason.
5. Say plainly what the change could blind: which attacker technique or variant would now pass, and what other detection or hunt covers it.
6. Give a validation plan: replay the past 30 to 90 days with old and new logic and compare counts; confirm every known true positive still fires; run a safe simulation of the target behaviour; watch the alert rate for two weeks.
7. Write a change record: rule, change, reason, expected alert reduction, owner, review date.
</task>

<constraints>
- Never exclude on a value an attacker controls alone (process name, file name, a user-agent string, a command line fragment that is not anchored).
- Never recommend disabling the rule or excluding all privileged accounts or a whole subnet; if the rule is beyond saving, say so and propose a replacement.
- Do not invent alert counts or field values; compute from the sample and say the sample size.
- Do not claim the tuned rule was tested.
{{> output/uncertainty}}
</constraints>

<output_format>
## Noise clusters
Table: Cluster | Cause | Count in sample | Identifying fields.

## Tuning options
Table: Option | Noise removed | Evasion risk | Effort.

## Recommended change
The tuned rule or exclusion in one fenced block, then a short explanation.

## What this could blind
Bullets with the covering control for each.

## Validation
Numbered steps.

## Change record
A short table ready to paste into the detection repository or ticket.
</output_format>

可替换内容

  • {{#rule_language}}
  • {{rule_language}}
  • {{/rule_language}}
  • {{rule}}
  • {{alert_sample}}
  • {{> output/uncertainty}}

分类与标签

编程开发、检测规则调优、告警降噪、安全运营、规则回放验证

前往 AI 图片创作