Home › Topics › Spotting Data-Exfiltration Patterns › Investigation Playbook

Investigating a Suspected Exfiltration: A Calm Playbook


The morning report has done its job. A finance workstation sent several gigabytes overnight to an address nobody on your network has ever contacted. It used a port it never uses, while its user was on leave. Three signals lined up, and the question is no longer "would we know?" but "what do we do?" For most small IT teams, nobody has written that down. So people improvise: pull the cable, email the user's manager, or stare at the graph for a week hoping it resolves itself.


Improvisation fails in two opposite directions. Panic destroys evidence, tips off whoever is on the other end, and can turn a contained problem into an uncontained one. Paralysis lets the transfer finish while you watch. A playbook exists so that at three in the morning, or at nine with a manager behind you, you follow steps rather than instincts. Calm is not a personality trait here; it is a procedure you wrote in advance.


This guide is that procedure, written for the defender who is usually also the network admin, the server admin, and the help desk. Seven steps: confirm, identify, preserve, assess, contain, escalate, communicate. It faces the hardest decision honestly — keep watching to learn the scope, or stop the bleed now. It ends with a one-page checklist and a decision matrix to print before you need them.


In this guide

  • Step 1: Confirm the pattern
  • Step 2: Identify the host, the user, and the destination
  • Step 3: Preserve evidence before touching anything
  • Step 4: Assess what that host could reach
  • Step 5: Contain proportionately, and who decides
  • Step 6: Escalate to the right people
  • Step 7: Communicate carefully, and document everything
  • The playbook on one page
  • Where to go next

Step 1: Confirm the pattern


The first question is not "who did this?" It is "is this real?" Monitoring systems produce false alarms for boring reasons. There might be a collector that double-counted a night's flows, a time-zone slip, or a wrong baseline row. Or a host might have been reassigned to a new role without anyone updating the inventory. Spend the first minutes ruling those out, because everything after this step gets more expensive and more visible. Confirming before reacting applies to every incident — The First Fifteen Minutes lays it out for outages; here it is applied to a suspected leak.


Confirm with a second source. If the alert came from flow data, look for the same sessions in the firewall log, and the other way round. Check the DNS log for the destination's name, or its absence. Re-derive the volume from raw records rather than trusting the summary. Then answer three questions, writing down each answer and the time:


  • Is it real? Two independent sources agree on host, destination, window, and roughly the volume.
  • Is it one-off or ongoing? Look back through the history for this host-and-destination pair. A single burst last night is one situation; the same destination every night for three weeks is a very different one.
  • Is it happening right now? An active transfer changes the urgency of everything that follows. Check the live session table on the firewall or the most recent flows.

If the pattern evaporates under a second source — a reporting artefact, a host that is actually the new backup proxy — close it, fix the cause, and note it. If it holds, you have a confirmed anomaly. Not a confirmed theft; the network cannot tell you that, and you should not say it. The five-dimension scoring in What Data Exfiltration Looks Like is a good structure for writing down exactly what is and is not abnormal.


Step 2: Identify the host, the user, and the destination


An IP address is not a host, a host is not a person, and a destination address is not a place. Step two turns each into something you can reason about, using lookups in data you already hold. The table below is the map.


You haveYou wantWhere to look
Internal IP address at a given timeDevice name, hardware address, switch portDHCP lease history for that time; the switch's address table; your asset inventory
Device nameRole, owner, location, what it holdsInventory, the baseline table, the desktop team's records
DeviceUser at the keyboard during the windowLogin and logoff events from the directory; VPN session logs for remote hosts; who physically sits there
UserStatus and contextDirectory attributes; leave or departure status from HR, asked through the right channel (see Step 6)
External IP addressThe name that was looked upDNS query log for the host, seconds before the session
External address or nameOwning network, rough reputationPublic registration data; your own known-good and known-neutral lists; whether any other host has ever contacted it

Two identification traps. First, the address may have moved. DHCP can hand the same address to a different device a day later. So resolve it against the lease active during the window, not the one active now. Second, the account may not be the person: a login tells you whose credentials were used, not whose hands were on the keyboard. Malware runs as the logged-in user; a stolen password is used by somebody else. Write "session under account j.doe," never "j.doe did this." That distinction protects an innocent colleague and keeps your case file accurate.


Also check whether the destination is only this host's problem. If three other hosts have started talking to the same place, you are no longer investigating one machine.


Step 3: Preserve evidence before touching anything


This step is where panic does its damage, so it gets the loudest instruction in the guide: preserve first, then act. Flow records age out, firewall logs rotate, DHCP lease history is overwritten. Logging into the host to "have a look" changes timestamps, may trigger anything watching for an administrator, and creates a story in which you altered the evidence. Every action from here on should be one you could describe, in order, to someone who was not there.


  • 1. Export the flow and firewall records for the host, for the alert window plus a generous margin. The whole day before and after is not excessive. Do not filter aggressively; you do not yet know what will matter.
  • 2. Export the DNS query log for the host over the same range, and the queries for the destination name from any host.
  • 3. Save the identification lookups from Step 2: the lease record, the login events, the switch address entry, as they were when you looked.
  • 4. Capture packet metadata if you can. If the transfer is still active and you have a capture point, headers alone are valuable even though the content is encrypted. Do it in a way that survives scrutiny later; Capture Hygiene covers what makes a capture evidence-grade rather than merely interesting.
  • 5. Note the time reference. Which clock each source uses, and whether it is local or UTC. A timeline built from mismatched clocks is a timeline that falls apart under questioning.
  • 6. Compute a hash of every export and record it with the file name, the time, and your name. Put the files somewhere write-once, or at least somewhere only you and your lead can reach.
  • 7. Start the investigator's log: a plain text file with a timestamped line for every action and every decision, beginning with the moment you saw the alert.

Twenty minutes of exporting now saves you from the sentence "we no longer have the records" later — the sentence that ends investigations, insurance claims, and sometimes careers.


Tip: agree a home for evidence exports before you need one. That could be a locked-down share or a dedicated small server with restricted access and no automatic cleanup. Deciding at three in the morning where to put files is how they end up on a desktop that gets reimaged.


Step 4: Assess what that host could reach


You know something left. You do not know what. But you can bound it, and the bound is what everyone above you will ask for first. The question is: what could this host, under that account, have reached? Not what it did reach — you usually cannot see that from the network — but the worst case.


Walk the access. Which shares are mapped on that host, and what do they contain? What does the account have rights to across the directory, file servers, business applications, and cloud services? Does the host hold cached credentials or saved sessions for anything more sensitive than itself? If internal flow data is available, look for the gather: unusual reads from internal servers in the hours or days before the outbound burst. Suppose a workstation pulled far more than usual from the contracts share on Wednesday afternoon. It then uploaded a similar amount to a first-seen address on Thursday morning. That tells you where to look.


Compare volumes. If the host sent four gigabytes and the most sensitive thing it can reach is a two-hundred-megabyte spreadsheet, the worst case is bounded. If the contracts share is three and a half gigabytes, it is not. Write the assessment as one sentence a manager can repeat: "In the worst case, this host could have exposed the finance share and the HR reports folder; it could not have reached the customer database." That sentence drives every decision in the next two steps.


Step 5: Contain proportionately, and who decides


Containment is a ladder; climb only as far as the situation demands. Each rung stops more, disrupts more, and is more visible to whoever is on the other end.


The containment ladder

  • Block the destination at the firewall. Cheapest, least disruptive, reversible; it stops this channel and leaves the host and its user untouched. The cost: whoever is on the other end will notice, and malware may simply try somewhere else.
  • Isolate the host: a quarantine VLAN, a shut switch port, or endpoint isolation if you have it. It stops every channel from that host, preserves its state for examination, and is loud — a present user will know within minutes.
  • Disable the account and reset its credentials, including saved sessions. Necessary when the account itself may be compromised or the concern is a person rather than a program. This rung touches a human being's employment and is never yours alone to climb.
  • Widen if Step 2 showed other hosts talking to the same destination: block, then isolate each in turn, preserving evidence first.

Who decides

Decide the authority in daylight, write it into the playbook, and do not renegotiate it during an incident. A reasonable split for a small team starts here: the on-call admin may block a destination immediately on their own judgement. That is because blocking a destination is cheap and reversible. Isolating a host needs a second person — the IT lead or whoever holds that role at night. That is because isolation disrupts someone's work and starts a visible event. Disabling a person's account needs management and, for an employee, HR; it is never an IT-only decision. Pulling the internet link for the whole organization is a management call, full stop.


Watch longer, or stop the bleed?

This is the tension at the heart of every exfiltration case. Stop now, and you limit the loss but learn less. The destination goes quiet, other involved hosts stay hidden, and you may never know the full scope. Watch longer, and you learn the scope — but every minute is more data gone. There is no universally correct answer; there is a way to decide. And it turns on two things you already established: is it still active (Step 1) and how sensitive is the reach (Step 4).


Low sensitivity reachHigh sensitivity reach
Active nowWatch briefly with a hard deadline (an hour is plenty) while preserving; then block the destination and reassess.Stop the bleed now: block the destination and isolate the host. Learn scope afterwards from the preserved records.
Finished, one-offPreserve, block the destination, investigate at working pace. Escalate at the normal tier.Preserve, block, isolate the host so it cannot resume, escalate immediately. Treat the data as exposed until proven otherwise.
Finished, but recurring nightlyPreserve; consider one more night of watching with a capture in place, then block before the next window.Block and isolate before the next window; do not trade another night for scope.

Two rules sit above the matrix. Never watch without preserving; watching is only worth anything if the records survive. And if you watch, write down what you are waiting to learn and when you will stop. "We are watching until we see whether a second host contacts the destination, or for one hour, whichever comes first" is a decision. "Let's see what happens" is not.


Step 6: Escalate to the right people


You are the network's witness, not its judge; the decisions that follow a confirmed anomaly belong to people with the authority to make them. Escalate early rather than late. Nobody is criticized for raising a confirmed pattern with a clear summary, and many people have been criticized for sitting on one. Who, roughly in order:


  • Your own management first, with the one-sentence worst case from Step 4, what you have already done, and what decision you need from them. Keep it short and factual; resist speculation about who or why.
  • Security or incident-response help, if you have any: an internal security person, a retained external responder, or the provider named in your cyber-insurance policy. If you have none, say so to management now; the decision to bring in outside help is theirs, and it is easier to make early.
  • Legal and HR the moment the picture involves a person rather than only a program. Investigating an employee has rules — about privacy, about evidence, about what you may and may not do. Those rules vary by place and by organization. The right move is to stop reasoning about the person and hand the human side to the people whose job it is.
  • Insurers and regulators are concepts here rather than your call. Many cyber-insurance policies require notification of a suspected incident within a set period. Many kinds of data carry legal notification obligations if exposed. You do not decide whether these apply. Your job is to make sure management and legal know a potential exposure exists in time for them to decide.

The escalation message is a template, not an essay. State what was observed, from which sources, and when. State what the host is and could reach, and what has been done and preserved. Then state what decision is needed now, and from whom. Four parts, and the people receiving it can act.


Step 7: Communicate carefully, and document everything


Communication during a suspected exfiltration is on a need-to-know basis, and the circle is smaller than instinct suggests. Do not post in the all-hands chat. Do not email the user's team asking whether anyone knows why the machine was busy. Do not discuss it with the user until HR and management have decided how and whether to. If the concern is an intruder rather than an insider, they may be reading the email system you are about to use to discuss them. In that case, agree an out-of-band channel — phone, a personal messaging app, a room — for the core group until the picture is clear.


Name a single point of contact for status; everyone else who asks gets "there is an ongoing investigation; updates come from this person at these times." Brief scheduled updates — even "no change" — stop the rumour mill that fills any silence. And choose words with care. "A confirmed unusual transfer from a workstation, under investigation" is accurate. "A data breach" and "an employee stealing data" are conclusions you do not have. They can cause harm if they turn out to be wrong.


Documenting the case

The investigator's log you started in Step 3 grows into the case file. It contains the timeline in one time reference, and the list of evidence exports with hashes and locations. It records every action taken, by whom, when, and on whose authority. It records every decision, including the decision to watch and its deadline. It also contains the worst-case assessment, who was notified and when, and the open questions still unanswered. It is boring to write and the most valuable thing you will produce. It supports an insurance claim, satisfies a regulator, protects an innocent colleague, protects you, and becomes the lessons-learned review. Write it as you go; memory at the end of a long day is not evidence.


The playbook on one page


The whole sequence as a checklist to print and pin up. The "who" column is a small-team default; adjust it to your own agreed authority.


StepDoOutputWho
1. ConfirmSecond source; one-off or ongoing; active now?Confirmed anomaly (or closed false alarm)On-call admin
2. IdentifyAddress to device; device to account; destination to name and owner; other hosts?Named host, role, reach; account; destination profileOn-call admin
3. PreserveExport flows, firewall and DNS logs, lookups; hash; note clocks; start the logEvidence set in a safe locationOn-call admin
4. AssessWalk the host's and account's access; look for the gather; compare volumesOne-sentence worst caseAdmin with the lead
5. ContainApply the decision matrix; climb the ladder only as far as neededDestination blocked / host isolated / account disabled, each recordedBlock: admin. Isolate: lead. Account: management and HR
6. EscalateManagement, security help, legal and HR if a person is involved; raise insurer and regulator questionsDecisions made by those with authorityLead and management
7. CommunicateNeed-to-know; single contact; scheduled updates; careful words; out-of-band if neededControlled information flow; growing case fileLead; case file kept by admin

Run through it once, on a quiet afternoon, against an imaginary alert. Fill in the real names for "lead" and "management," the evidence location, and the phone numbers. The first time you use the playbook should not be the first time you read it.


Where to go next


You now have a procedure that replaces instinct. Confirm before you react, identify without accusing, and preserve before you touch. Bound the worst case and contain only as far as needed with authority agreed in advance. Escalate to the people whose decisions these are, and communicate to a small circle while writing everything down. The steps that feed this playbook are the signals, the baselines, the after-hours rules, and the reports that raise the alert in the first place. They are the rest of Spotting Data-Exfiltration Patterns. Start with Outbound Baselines and Egress Monitoring on a Budget. And the broader craft of running any incident without panic is the subject of the Incident Response guides.


Put this into practice with SysPeep Monitor

Step 1 of this playbook is a second source, and Step 4 is knowing what a host is and what it can reach. SysPeep Monitor keeps both close at hand. It auto-discovers devices and interfaces so the inventory behind an address is already there. It graphs interface traffic over SNMP so the night's spike is visible beside weeks of normal. It shows it all on a live dashboard and network map from one web console. The free trial installs on any Windows machine.

Download Free TrialLearn About SysPeep