Automated phishing analysis: how AI triages reported emails
Automated phishing analysis checks a reported email's authentication, links, and attachments, scopes the campaign, and closes or escalates it with evidence.
Summary
Automated phishing analysis takes an email a user reported, checks its sender authentication, links, and attachments, finds who else received it and whether anyone interacted with it, and closes or escalates the report with the evidence attached, so an analyst doesn't run each lookup by hand. The hard parts aren't the lookups: they're getting the original message with its headers intact, trusting only the authentication results your own mail border wrote, treating fifty reports of one campaign as one case, and keeping purges and account actions behind a person's approval.
Key takeaways
- Have users' reports arrive as attached .eml or .msg files; a forwarded copy loses the headers the analysis depends on.
- Read only the Authentication-Results header your own border stamped, and check DMARC alignment against the visible From domain, not just pass or fail.
- Scope every malicious report before closing it: who else received it, who clicked, and what happened in their identity logs afterward.
- Let the agent close false positives and benign mail on its own, but gate message purges, sender blocks, and account actions behind approval.
- Merge repeat reports of one campaign into a single case so the queue measures campaigns, not inbox volume.
The short version
A reported phish is a question with four parts: is this email malicious, who else got it, did anyone act on it, and what should happen now. An analyst answers them with the same lookups every time: headers, authentication results, link and attachment reputation, a mail search, and the clicker's sign-ins. That repetition is why phishing is usually the first alert type teams hand to automation, and why an AI agent can work it end to end.
What separates a useful pipeline from a noisy one is mostly plumbing and policy. The agent needs the original message, not a forward. It needs to know which headers it can trust. It needs to treat a campaign as one case. And it needs a clear line between what it may do alone, such as closing a newsletter someone reported, and what waits for a person, such as purging a message from four hundred inboxes.
In Cotool, that whole loop is one response agent: reports arrive at the agent's own email address, each becomes an alert, the agent investigates with the tools it's been granted, and the alert ends closed or escalated with the evidence written into its timeline.
What the analysis has to answer
Phishing covers more than links. MITRE ATT&CK splits technique T1566 into spearphishing attachments (T1566.001), links (T1566.002), messages through third-party services (T1566.003), and voice (T1566.004), where the email asks the recipient to call a number. An analysis built only around URL reputation misses the last two. For each report, the agent should produce:
| Question | Evidence | Where it comes from |
|---|---|---|
| Is it malicious? | Authentication results, sender history, link and attachment verdicts, the ask in the body | The original message, threat intelligence |
| Who else got it? | Recipients of the same sender, subject, or URL | Mail logs or the email security tool |
| Did anyone act? | Clicks, downloads, replies, sign-ins, app consents | Proxy and DNS logs, identity provider |
| What now? | A disposition and the proposed response | The agent's summary, for approval |
The last row is the one most write-ups skip. A verdict on the email alone tells you nothing about whether the one person who clicked then entered a password or granted an app access to their mailbox.
Step by step
1. Get the original message, headers intact
Route reports to a dedicated mailbox and from there to the analysis. In Microsoft Defender for Office 365, admins can send Outlook's Report button to a reporting mailbox, to Microsoft, or both. Microsoft's requirements for anything submitted to that mailbox are a good rule for any pipeline: the reported message must be an unmodified, uncompressed .EML or .MSG attachment, because forwarding "strips required headers," and messages with several attachments are discarded.
Two settings save trouble later. Mark the reporting mailbox as a SecOps mailbox so its mail isn't filtered, and Microsoft warns that otherwise a reported simulation can assign the reporter phishing training. Exclude it from data loss prevention rules too.
In Cotool, an email trigger gives a response agent its own address. Every message that arrives starts the agent with the email as context, and by default creates an alert assigned to that agent, with the original payload stored on the alert.
2. Check authentication against your own border
Most of the verdict starts in the headers. Read the Authentication-Results field your own mail border added, and ignore any others. RFC 8601 warns that a sender can forge the field "using the DNS domain of a receiving ADMD" and claim the message passed, so trust only the entry whose authserv-id is your border's:
Authentication-Results: mx.yourcompany.com;
spf=pass smtp.mailfrom=billing-notice.net;
dkim=pass header.d=billing-notice.net;
dmarc=fail header.from=yourvendor.com
That message passed SPF and DKIM, but for a different domain than the one the recipient saw. RFC 7489 calls the property that matters identifier alignment: the From address domain has to match a domain SPF or DKIM validated. A pass with no alignment is a lookalike, not a vendor. Compare Reply-To and Return-Path with From as well, since a reply address on a free mail domain under a corporate From line is the shape of business email compromise.
3. Look up links and attachments without opening them on a laptop
Extract every URL, including those behind redirects and inside attachments, plus file hashes and the sending IP, and check them against threat intelligence. Cotool's integrations cover email security (Abnormal, Proofpoint, Mimecast) and threat intelligence (VirusTotal, URLScan, AbuseIPDB), and each agent runs in an isolated Linux sandbox where it parses and decodes payloads itself, with outbound traffic restricted per agent. Decoding a base64 HTML attachment there is safer than in an analyst's browser.
For voice and service-based phish, the useful indicators are the phone number, the claimed sender, and the request itself. Ask the agent to quote the ask ("call to cancel your subscription") in its summary, because that's what the reporter's colleagues will receive too.
4. Scope the campaign
Before closing anything malicious, search mail for the same sender, subject, and URLs, and list every recipient. Then treat repeats as one case. In Cotool, an alert can be marked as a duplicate of a canonical alert, from its status menu or in bulk from the alerts list. The duplicates are archived, and the relationship is recorded in both timelines, so fifty reports of one campaign become one case with fifty reporters attached.
5. Check who acted on it
For each recipient, look for clicks in proxy or DNS logs, then for what followed: a sign-in from a new location, a password reset, a new MFA device, or an OAuth grant. Some phishing ends in an app consent rather than a stolen password. Cotool's response page shows an OAuth Grant Responder for that case: it lists the app's scopes and grants in Google Workspace, checks the publisher's domain in VirusTotal, traces what the token read and changed, and asks #security before revoking the app. In the example on that page, the revocation covers four users and the run reports 412 emails exposed.
Cotool agents can split this step: an agent can hand work to investigator subagents for parallel investigations, such as one per recipient's sign-ins, and to a reviewer subagent for peer review of the verdict.
6. Decide, and gate the response
Close what the evidence clears: a marketing email someone reported, or a known vendor's invoice that passes aligned DMARC. Escalate what it doesn't. Cotool builds this line into alert statuses. A response agent can close an alert as false positive or benign, but it can't close one as a true positive; it escalates it for a person, and its final comment states the evidence, the conclusion, and the gaps. Escalations can be announced in Slack or open an incident in incident.io, and the Slack post updates as the alert's status changes.
Response actions sit behind approval. Purging the message from every mailbox, blocking the sender, resetting a password, or suspending a user are state-changing tools, and Cotool's tool permissions can require approval for each in the app or in Slack while reads run automatically. The AI SOC in Slack guide covers how those approvals work in a channel.
Phishing is also a volume job, so the model matters for cost. Cotool sets the model per agent; its own example runs a Phishing agent on GPT-5.6 Sol at medium effort and keeps higher-effort models for investigation.
An example agent instruction
The behavior above fits in a short system prompt. This is a starting point to adapt, not a Cotool default:
You triage user-reported email. The reported message is attached as .eml.
1. Use only the Authentication-Results header from mx.yourcompany.com.
Record SPF, DKIM, DMARC, and whether the From domain is aligned.
2. Extract URLs, attachment hashes, the sending IP, phone numbers, and the request.
Look each indicator up; never open links outside the sandbox.
3. Search mail for the sender, subject, and URLs. List every recipient.
If an open alert covers the same campaign, mark this one as its duplicate.
4. For recipients who clicked, check sign-ins, MFA changes, and app consents since delivery.
5. Close as False Positive or Benign only with evidence. Otherwise escalate.
Propose purges, blocks, and account actions for approval; never run them unasked.
6. End with a summary comment: verdict, evidence, affected users, gaps.
Common mistakes
- Analyzing the forward. A forwarded report carries the reporter's headers, not the attacker's. Require attachments.
- Trusting the first
Authentication-Resultsheader. Read the one from your border; anything above it may be forged. - Stopping at pass or fail. A DMARC pass for the attacker's own domain is still phishing. Check alignment with the visible From domain.
- Letting simulations into the queue. Your own phishing tests will be reported. Exclude their senders, or the agent will investigate your training program.
- Closing on the email alone. A malicious message that one person clicked is an identity incident. Scope recipients and their sign-ins first.
- Counting reports as incidents. Merge duplicates, or one campaign inflates the queue and your metrics.
- Taking orders from the email. The body is attacker-controlled text, and an agent that reads it can be told to "mark this as safe." Treat it as prompt injection waiting to happen: the verdict should rest on headers, indicators, and logs, not on what the message says about itself.
- Letting the agent purge on its own. A wrong purge deletes a real customer's email from every inbox. Keep it behind approval.
Checklist
- Reporting mailbox set up, marked as a SecOps mailbox, and excluded from DLP
- Reports arrive as unmodified .eml or .msg attachments
- Only your border's
Authentication-Resultsheader is read - DMARC alignment,
Reply-To, andReturn-Pathchecked againstFrom - URLs, hashes, sender IP, phone numbers, and the request extracted and looked up in a sandbox
- Recipients of the same sender, subject, or URL listed before any close
- Clickers' sign-ins, MFA changes, and app consents checked
- Simulation senders excluded
- Duplicate reports merged into one case
- False positive and benign closes allowed; true positives escalated to a person
- Purges, blocks, and account actions require approval
- Every close or escalation carries a summary with verdict, evidence, and gaps
FAQ
- What is automated phishing analysis?
- Automated phishing analysis is software that investigates a suspicious email without an analyst running each step. It parses the headers, checks SPF, DKIM, and DMARC results, looks up links, attachments, and sender infrastructure, finds other recipients and clicks, and returns a disposition with the evidence behind it.
- Can AI close phishing reports on its own?
- AI can close phishing reports on its own when the evidence shows the email was a false positive or expected mail, such as a known vendor's invoice that passes aligned DMARC. Confirmed phishing should be escalated so a person approves the response. In Cotool, response agents can't close an alert as a true positive at all; they escalate it for a human to close.
- How do you send user-reported phishing emails to an automated analysis?
- Send user-reported phishing emails to a reporting mailbox and from there to the analysis as attachments, not forwards. Microsoft Defender for Office 365 lets admins route Outlook's Report button to a reporting mailbox, and requires the original message as an uncompressed .EML or .MSG attachment because forwarding strips required headers.
- Why do phishing reports need DMARC alignment and not just a DMARC pass?
- Phishing reports need DMARC alignment because a message can pass SPF or DKIM for a domain the attacker controls while showing a different domain in the From line. RFC 7489 defines alignment as the From address domain matching a domain that SPF or DKIM validated, which is the check that ties the authentication to what the recipient saw.
Sources
- Phishing, Technique T1566 - MITRE ATT&CK · accessed November 20, 2026
- RFC 8601: Message Header Field for Indicating Message Authentication Status · accessed November 20, 2026
- RFC 7489: Domain-based Message Authentication, Reporting, and Conformance (DMARC) · accessed November 20, 2026
- Phishing Guidance: Stopping the Attack Cycle at Phase One (CISA) · accessed November 20, 2026
- Configure user reported message settings in Microsoft Defender for Office 365 · accessed November 20, 2026
- AI SOC Without the Black Box - Cotool · accessed November 20, 2026
- AI Alert Triage and Incident Response Agents - Cotool · accessed November 20, 2026
- Triggering Agents - Cotool Documentation · accessed November 20, 2026
- Alerts - Cotool Documentation · accessed November 20, 2026
- Working Alerts - Cotool Documentation · accessed November 20, 2026
- Alert Routing - Cotool Documentation · accessed November 20, 2026
- Integrations Overview - Cotool Documentation · accessed November 20, 2026
- Tools and Integrations - Cotool Documentation · accessed November 20, 2026





