调查异常登录与账号入侵告警
结合登录时间线、多因素认证及后续操作核验告警,给出附置信度的判断、处置步骤与补查清单。
完整提示词
<context>
Identity alerts are the noisiest and the most important alerts a SOC handles. Most impossible-travel alerts are corporate VPN or proxy egress, cloud-hosted browsers, mobile carriers whose IPs geolocate to another city, or a phone and a laptop on different networks. The real attacks look different: a burst of MFA prompts the user did not start, followed by one approval; a session from a hosting provider or residential proxy that never did MFA because it replayed a stolen token; a successful sign-in followed within minutes by a new MFA method, an inbox rule, an OAuth consent or a password change. The question is not "is the location odd" but "did someone other than the user get a working session, and what did they do with it".
</context>
<task>
Investigate this {{identity_provider}} sign-in alert.
<sign_in_data>
{{sign_in_data}}
</sign_in_data>
{{#user_context}}
<user_context>
{{user_context}}
</user_context>
{{/user_context}}
1. Summarise what the data shows as a short timeline: each sign-in with IP, network owner type (corporate, residential ISP, mobile carrier, hosting or VPN provider, if the data shows it), location, device and client, MFA outcome, and what happened next.
2. Test each benign explanation against the data and say whether it fits: corporate VPN or proxy, travel, mobile carrier geolocation, new device, a service or sync client, a password manager or browser extension. Name the field that supports or rules out each one.
3. Look for the attack patterns: repeated MFA denials or timeouts before an approval; single-factor sign-ins that reuse an existing token; legacy protocols that skip MFA; two active sessions from distant places at once; post-sign-in changes (MFA method, inbox rule, forwarding, app consent, password, privileged role).
4. Give a verdict: true positive, benign true positive (real event, expected), false positive, or undetermined, with a confidence level and the two or three facts that decide it.
5. Actions for the verdict. For a true positive or undetermined with risk: revoke sessions and tokens, reset credentials, remove unknown MFA methods, block the source, check the same IP or user agent across other users, and contact the user out of band (not by email to the possibly compromised mailbox). For a false positive: the tuning or named-location change that would stop the repeat, without excluding the user entirely.
6. List what to check next if anything remains undetermined.
</task>
<constraints>
- Never decide on location alone. Every verdict cites at least two facts from the data.
- Do not invent ASN owners, geolocation or user travel. If the network owner is not in the data, say it must be looked up.
- Do not suggest excluding a user, an admin or a whole country from the alert as a fix.
- Asking a user "was this you?" is useful but not proof; say how much weight their answer gets.
{{> output/uncertainty}}
</constraints>
<output_format>
## What the data shows
Table: Time (UTC) | IP and network type | Location | Device or client | MFA | What followed.
## Benign explanations tested
Table: Explanation | Fits? | Deciding field.
## Verdict
One line, confidence, then the deciding facts.
## Actions
Numbered, each with owner role.
## Check next
Bullets, or "Nothing outstanding".
</output_format>可替换内容
- {{identity_provider}}
- {{sign_in_data}}
- {{#user_context}}
- {{user_context}}
- {{/user_context}}
- {{> output/uncertainty}}
分类与标签
编程开发、异常登录、账号安全、多因素认证、安全事件响应