Short · Fraud and phishing · 18 min read

How Should a SOC Triage Employee-Reported Phishing Emails?

Treat an employee's report as an intake event that must become an evidence-backed decision, not as an incident verdict or a completed response. Preserve the reported message and relevant delivery context, determine whether anyone has already acted on it, correlate related copies without discarding meaningful differences, and verify the scope and outcome of any response. Close the report only when the reason for closure is recorded and unresolved exposure has an owner.

A report button can make it easier to ask for help. It cannot establish that the message reached the intended team, that a vendor analyzed it, that related messages were found, or that a removal action succeeded. Those are separate parts of the service. A healthy workflow makes their boundaries visible instead of turning them into one green status.

I would organize the work around two questions: what do we know about this message, and what do we need to do about the affected people and systems? The first supports classification. The second supports incident response. They often progress together, but waiting for a final classification should not prevent an appropriate, bounded response to credible evidence of active harm.

Define the service before measuring its queue

This paper addresses suspicious email reported by employees to a security operations team. It develops an operational framework from twelve original standards and official publications. Microsoft 365 provides concrete implementation examples; equivalent capabilities and limitations must be established separately for other platforms. The proposed framework and fictional example are analytical synthesis, not results from a deployed service or an experiment.

The service begins when an employee uses a recognized reporting route. Its output should be a reasoned disposition, any necessary protective action and useful feedback. The disposition might be confirmed malicious, legitimate, unwanted but nonmalicious, insufficient evidence, or transferred to a broader investigation. These are proposed categories, not a mandated industry taxonomy.

A report is not interchangeable with a message, a recipient exposure or a campaign. Several employees may report copies of one malicious message. One employee may report the same copy twice. A campaign may contain many variants, and a single report may reveal account compromise extending beyond email. The workflow needs relationships among those objects without pretending that they are identical.

NIST recommends prioritizing incidents according to scope, likely impact, urgency and available resources, and tracking their status as circumstances change (Nelson et al., 2025, RS.MA-03–04). I would apply those principles to reported-email triage by giving every report an accountable handling state, while reserving incident escalation for the evidence and consequences that warrant it.

Verify the reporting route end to end

Start with the path the employee actually uses. Does the action send a message to a monitored organizational mailbox, to the platform provider, to both, or somewhere else? Does the route differ between desktop, web, mobile and shared-mailbox use? Those are deployment questions, and they should be checked rather than answered from the button's label.

Microsoft's user-reported settings distinguish reporting to Microsoft, to an organizational reporting mailbox, or to both. Its mailbox-only option does not submit the message to Microsoft unless an administrator does so subsequently (Microsoft, 2026e, “Use the Microsoft Defender portal to configure user reported settings”). Receipt in the local queue and vendor analysis are therefore different facts.

I would validate the route using approved harmless material. Record the reporting action, the destination, the resulting identifier and what the receiving team can actually inspect. Include a case in which an integration or downstream queue is unavailable. A confirmation displayed to the employee should mean something specific; it should not imply completed analysis if it only confirms submission.

Assign ownership for failures between systems. If a report reaches a mailbox but does not create a case, someone must detect and resolve that gap. If an external analysis service is unavailable, the local team needs an explicit fallback disposition. Otherwise a workflow can lose work while every individual component appears to be operating within its own narrow responsibility.

Reporting configuration is also a data-handling decision. A suspicious message may contain legitimate customer information, internal correspondence or attachments unrelated to the threat. Decide where reports may be sent and who may access them. The investigation should not silently distribute complete messages to additional services merely because a convenient integration exists.

Preserve an analyzable object, not only a screenshot

A screenshot can explain what worried the employee, but it is not equivalent to the original message. Retain the available native message object, relevant headers, attachments and delivery metadata using the approved collection method. Preserve the employee's description as a separate observation, including whether they clicked, opened, replied, entered information or approved a request.

The Internet Message Format distinguishes author, destination, trace and identification fields. Its Message-ID identifies a particular version of a message and is generated by the sending host; it is not a cryptographic statement about the sender's honesty (Resnick, 2008, §§3.6.2–3.6.4, 3.6.7). Treat header values as evidence with a provenance, not as unquestionable identity assertions.

MIME defines how message content is described and transfer-encoded, including mechanisms such as base64 and quoted-printable (Freed & Borenstein, 1996, §§5–6). The implication for triage is to retain both the acquired object and the controlled interpretation used for analysis. A rendered view may omit the representation needed to explain how a link, attachment or nested content was interpreted.

Record collection time, collector, source, format and transformations. Keep an integrity value for the acquired file and work from a controlled copy where appropriate. NIST's forensic guidance describes verifying file integrity during acquisition and examination through message digests (Kent et al., 2006, §4.2.2). Its historical algorithm examples are not a recommendation to select an obsolete algorithm today.

An integrity match establishes that the retained bytes have not changed relative to the recorded value. It does not establish that the initial export was complete, that the sending party was legitimate, or that every mailbox copy is identical. Those are additional claims requiring different evidence. This distinction keeps preservation from becoming a misleading badge of authenticity.

Give every identifier a defined scope

A case should retain enough identifiers to connect the employee's report to the provider's records and the original message. Do not use a subject line as the sole key. It is human-readable context, but many unrelated messages can share it, and related messages can vary it.

Microsoft's EmailEvents schema distinguishes NetworkMessageId, generated by Microsoft 365, from InternetMessageId, set by the sending system. It also includes recipient and delivery-related fields (Microsoft, 2026b). These fields should be used according to their documented meaning rather than collapsed into a generic message ID column.

I would store the report identifier, provider message identifier where available, internet Message-ID, recipient, relevant timestamps and acquisition reference separately. The same campaign can involve several message identifiers. Conversely, one message identifier in a transport system may need to be interpreted alongside recipient-specific delivery evidence. The investigation model should preserve that distinction.

Maintain a record of normalization. Lowercasing a domain for comparison, extracting a URL host or decoding an attachment can be useful analytical steps. Preserve the original value and the transformation used. An analyst should be able to explain why two artifacts were grouped and recover the inputs if the grouping later proves wrong.

Do not turn message content into a freely searchable organizational archive without a defined purpose. Analysts may need detailed material for a case, while service managers need aggregate queue evidence. Separate those access needs. A retention decision should consider investigation value and sensitivity rather than simply keeping every reported mailbox forever.

Authenticate the evidence about authentication

Authentication results can contribute to classification, but only when the analyst knows where the result came from. A string saying pass inside a message is not self-authenticating evidence from the organization's receiving system.

RFC 8601 describes forged Authentication-Results fields and the trust-boundary handling needed to prevent receivers from accepting claims inserted by an untrusted source. It also discusses authentication fields inside encapsulated messages (Kucherawy, 2019, §§5, 7.1, 7.10). When reviewing a forwarded report, distinguish the authentication of the reporting wrapper from the results associated with the suspicious original.

Even a trustworthy authentication pass has a limited meaning. The current DMARC specification explicitly states that a pass validates authorized use of the Author Domain but does not establish that delivery is safe or desirable (Herr & Levine, 2026, §1). An authenticated message can still contain a harmful request. Classification must consider the content, context and requested action.

I would record the relevant receiver's result, the identifier evaluated and the source of that evidence. If the original is unavailable or the report has been transformed, state the limit. Do not manufacture a confident conclusion by interpreting the report's own successful delivery as evidence about the earlier message.

The separate DMARC sending-estate assessment addresses whether legitimate organizational senders authenticate correctly. Here the task is different: deciding how to handle a reported message and its associated exposure. Domain-authentication deployment is one input, not the entire triage workflow.

Establish exposure before spending the whole shift on classification

Ask what happened, not merely whether the message looks suspicious. Was it blocked before delivery, delivered to a mailbox, opened, acted on or forwarded? Which of those states is supported by telemetry, which is reported by the employee, and which is unknown? Preserve the distinction between an observation and an inference.

An employee saying they entered a password into an unexpected page changes the immediate response question even if the vendor's message verdict is still pending. The team should follow its established account-response process within authorized scope. Removing the email alone would not establish that the resulting account exposure has been addressed.

Similarly, an attachment reportedly opened on a managed device raises a device-investigation question. A fraudulent payment instruction may require a business-process response. These are proposed escalation examples; they are not claims that every click causes compromise or that every suspicious invoice is fraud.

I would use a small set of explicit priority factors: evidence of completed harmful action, sensitivity of the affected account or workflow, delivery scope, time-sensitive business consequences and the confidence of the available evidence. Keep uncertainty visible. A single report with credible active exposure may deserve attention before hundreds of obvious duplicates already covered by a verified response.

Do not require the employee to perform additional risky actions to complete the report. The analyst should obtain missing evidence through approved systems where possible. Questions to the reporter should clarify actions already taken and business context, not ask them to reopen an attachment or visit a suspicious destination to reproduce the issue.

Correlate reports without erasing the exception

Grouping related reports can reduce duplicated analysis, but correlation is a hypothesis about shared meaning. Identical subjects, display names or URL hosts are useful starting points, not sufficient proof that every message deserves the same disposition.

I propose three layers. A report group joins repeated submissions concerning the same acquired or provider-identified object. A message-variant group joins copies supported as materially equivalent for the decision being made. A campaign group joins related variants while retaining their differences. These are analytical groupings, not claims that a vendor's interface implements precisely this model.

For each grouping rule, specify the evidence and the consequence. Exact attachment hashes may support grouping that attachment's content, but they do not establish that every surrounding message asks the same thing. A common domain may belong to a widely used service. Subject similarity can identify a theme while being a poor basis for destructive remediation.

Keep recipient-specific exposure and action status attached to the group. If one copy was quarantined and another remains delivered, the campaign is not uniformly contained. If one employee acted on the request, closing the message analysis should not close that person's incident automatically.

A new report should be checked for meaningful differences before inheriting a completed disposition. Changed attachments, destinations, authentication context, recipient actions or requested business transactions may reopen the analysis. Record why a report was treated as equivalent, and provide a way to split it back out. Efficient triage should reduce repeated work without suppressing new evidence.

Use vendor analysis as an input with its own lifecycle

A submission service has eligibility, availability and rate constraints. Those conditions affect what it can establish at a particular time. A failed or unavailable submission is not a verdict that the message is safe.

Microsoft's submission documentation says administrators can submit messages up to thirty days old when the messages remain available and have not been purged. It also documents throttling and explains when resubmitting an already analyzed user report is useful, such as disagreement with the original verdict (Microsoft, 2026d, “Prerequisites”). These are concrete reasons to preserve submission status separately from case status.

I would record what was submitted: the message, a URL or an attachment. A verdict about one artifact should not silently become a verdict about the complete conversation or the sender's account. Retain the analysis time and result so that later evidence can be compared with the earlier decision.

If a provider returns no threat found, reconcile that result with the reported behavior and the available evidence. An apparently ordinary cloud-hosted document can carry a fraudulent business request without presenting a conventional malicious attachment. Conversely, a user's suspicion can concern legitimate but unexpected mail. The task is to resolve the specific concern, not to make the employee or the tool infallible.

Define when to escalate disagreement, seek additional context or close with a bounded conclusion. “No threat found in the submitted attachment” is narrower than “the request is authorized.” The narrower statement may be exactly what the tool established and should be reported that way.

Separate a response request from a verified result

A response workflow should record four different things: the intended scope, the action requested, the platform's result and the evidence of the resulting state. A successful request does not necessarily mean every intended mailbox or copy was affected.

Microsoft's Email entity documentation states that Latest delivery location reflects system or administrator actions but does not include subsequent user actions such as moving or deleting the message. It therefore does not guarantee the message's current location (Microsoft, 2026a, “What's on the Email entity page”). That is a specific limitation in a field whose name can invite overconfidence.

Automated remediation also has scope limits. Microsoft documents that email ZAP operates on cloud mailboxes rather than protected on-premises mailboxes and searches the last forty-eight hours of delivered mail (Microsoft, 2026f, “Zero-hour auto purge (ZAP) for email messages”). A team should not infer that a recent automatic action searched every historical copy across a hybrid estate.

I would reconcile the affected-recipient inventory with the action results. Identify successful, failed, unsupported and still-unknown cases. If the product cannot expose a definitive current state, record the strongest supported conclusion and the remaining verification task. Do not fill the gap with a status copied from an earlier stage.

Mailbox remediation and incident remediation must also remain separate. Quarantining an email can reduce continued access to that copy. It does not reverse a completed transaction, remove an already downloaded file, or establish that credentials were not disclosed. The incident record should retain those downstream questions even after the email queue item is resolved.

Prefer bounded, reviewable changes over broad cleanup

A query that is useful for finding related mail may be too broad for removing it. Before a destructive action, inspect the actual candidate set and confirm which fields the selected tool supports for that operation. Search syntax and remediation semantics are part of the evidence, not incidental implementation details.

Microsoft's eDiscovery deletion guidance requires validating the search scope and reviewing an export report before purge. It distinguishes feature-dependent workflows and limits, warns about unsupported search conditions, and states that the documented PowerShell purge does not delete unindexed items (Microsoft, 2026c, “Before you begin” and “Frequently asked questions”). These details make a universal “search and delete everything matching the subject” instruction indefensible.

I would choose the least disruptive action capable of meeting the incident objective, with authority and scope explicit. Where a reversible protective action is appropriate and supported, it can preserve options while analysis continues. Where removal is necessary, preserve required evidence first and verify the operation's coverage afterward. Existing retention or investigation obligations should be handled through the organization's established process; this paper does not prescribe changes to them.

Keep future-delivery controls separate from treatment of existing messages. Blocking a domain or changing a filtering rule can affect legitimate correspondence beyond the reported set. The review needs to explain that prospective effect, its owner and how errors will be recognized. One malicious message from a shared service does not establish that every message from that service should be blocked.

A false positive needs a recovery path too. Record how mistakenly restricted legitimate mail will be restored or released where the platform permits, how affected users will be informed and how the decision rule will be corrected. The ability to act quickly is incomplete without a way to detect and repair an overbroad action.

Give the reporter an answer that matches the evidence

The employee should receive an acknowledgment that confirms the report reached the service, followed by an appropriate disposition when available. These are different communications. An acknowledgment should not tell someone an email is safe before analysis has occurred.

A useful closing response identifies the conclusion and the next action without exposing unnecessary investigation detail. For example, it may confirm that the reported copy was handled and ask the employee to disclose any action already taken through the approved channel. If the evidence is incomplete, say so. Do not turn uncertainty into reassurance merely to close the ticket.

Make the response consistent with the case. If the case has transferred to an account investigation, the reporter should not receive a template suggesting that deleting the email finished the matter. If the message is legitimate, explain the relevant business context where appropriate instead of treating the report as a mistake deserving blame.

I would retain an explicit reason for every terminal state: decision supported, report merged with a named case, incident transferred with ownership accepted, or evidence insufficient with a defined next step. Merging should preserve a link to the group and its outcome. It should not make the original report disappear from accountability.

Measure service outcomes without rewarding superficial closure

Queue volume measures demand on the reporting service, not the number of attacks prevented. The same event can produce many reports, and a very consequential event can produce only one. Reporting totals should remain separate from unique message variants, campaign hypotheses and affected recipients.

I would retain timestamps for receipt, first substantive assessment, protective-action request and verified response. Define each transition before calculating elapsed time. Automated acknowledgment is not substantive assessment; an action submitted is not an action verified. Report waiting reasons separately so that an unavailable integration is not confused with an analyst's decision delay.

A small operational review can examine unresolved high-consequence exposure, reports lost between systems, incorrect grouping, actions with incomplete coverage and legitimate messages affected by response. These are proposed service measures, not validated benchmarks. Their value lies in identifying a fixable failure rather than producing an attractive aggregate number.

Measure workload at the group and exception levels as well as the report level. Efficient correlation may reduce repeated analysis while leaving a difficult minority requiring careful investigation. A falling average handling time can coexist with an aging set of unresolved cases. Review those cases directly instead of letting quick duplicate closures obscure them.

For the broader question of evaluating detection and response operations, assessing whether a SOC is effective provides the adjacent framework. This paper's narrower unit is the reported-email service and the decisions it must complete.

Worked example: one campaign, three different outcomes

Consider fictional Alderwick, whose employees report a set of invoice-themed messages. The first several reports concern the same attachment and request. The team preserves a representative original, links the related provider records and records recipient-specific delivery evidence. This is an invented example, not a client engagement or an observed incident.

The SOC confirms a harmful attachment and requests an approved protective action for the verified variant. Most results indicate the intended action succeeded, but several are unresolved. The analyst records those exceptions instead of marking the entire group complete. The case remains open for response verification even though classification of the attachment is complete.

A later report has the same subject but a different attachment and an ordinary business explanation. The grouping rule flags the difference, and the analyst separates the report. This protects legitimate correspondence from inheriting a destructive action based only on subject similarity. The earlier investigation remains useful context, not an automatic verdict.

Another employee reports that they already opened the harmful attachment. That report belongs to the original message group, but its exposure requires a separate device-response task. Removing the remaining mailbox copy cannot answer whether execution occurred or what happened afterward. Ownership of the device investigation is recorded before the report is closed as transferred.

The example demonstrates why classification, correlation, response coverage and downstream incident handling need distinct states. Collapsing them would produce a simpler dashboard with a weaker account of the actual outcome. Keeping them separate allows the team to reuse analysis while preserving the exceptions that change the decision.

Limits, tradeoffs and the first assessment to run

One objection is that detailed evidence handling will slow a high-volume queue. The answer is proportionate collection and reusable analysis, not abandoning provenance. A routine report need not become a full forensic acquisition, but the team still needs enough evidence to explain what it reviewed and why it acted.

Another objection is that automation should close obvious duplicates immediately. Automation can support bounded equivalence rules, provided the workflow retains exceptions and can reopen a group when new evidence appears. The cost of review should be weighed against the consequence of inheriting the wrong verdict or action. No grouping method is established here as universally optimal.

The framework and fictional example have not been empirically validated. Protocol documents define message and authentication semantics; vendor documentation describes particular capabilities and limits. Neither supplies a controlled comparison demonstrating that this workflow reduces incidents by a specified amount. Licensing, retention, client behavior and integration design require verification in the actual deployment.

I would start with an approved harmless report and trace it through intake, evidence collection, grouping, analysis and disposition. Then exercise a missing-original case, a duplicate with a meaningful difference, an unavailable integration and a partially successful response using a safe test design. The objective is to locate where the service loses evidence, authority or accountability before a real case depends on that transition.

Discuss a reported-email triage assessment if your report button works but the downstream result is unclear. A useful engagement would map the actual reporting routes, define evidence and decision states, review correlation rules and establish response-verification criteria. The result should show what the team can conclude, what it can do and how it will know the intended action happened.

References