Contribute a skill

Data Exfiltration Investigation

description

Investigate an escalated data-exfiltration alert. Quantifies what actually left — volume against baseline, the data's sensitivity, and the destination — attributes the transfer to a principal and channel, and distinguishes a backup or sanctioned sync from theft before returning a verdict with scope and containment actions.

Data Exfiltration Investigation

Runs on alerts escalated by triage (see the cloud and application exfiltration triage skills): a principal moved an unusual volume of data, or sent it to an unusual destination, in a way triage judged anomalous. Triage established the transfer looked wrong; investigation quantifies it and attributes it. The question is not "did data move" — it did — but "how much, how sensitive, where to, and who sent it."

Inputs

  • The escalated alert (principal, source store, destination, volume, timing) and the triage report.
  • Access and data-transfer logs (cloud storage access logs, DLP events, proxy/egress logs, SaaS download/export logs), the principal's activity history, and data-classification metadata where available.

Investigation steps

Collect all four signals; the verdict is derived from them at the end.

1. Volume and baseline — signal A

Measure what actually moved and compare it to the principal's own history and their peers'. Signal A fires when the volume is a sharp departure from baseline — orders of magnitude over what this principal, role, or service normally transfers — and especially a bulk pull (a whole bucket, a full mailbox, an entire table) rather than incremental access.

2. Sensitivity — signal B

Classify what left, not just how much. Signal B fires when the data is sensitive — credentials/secrets, regulated PII/PHI, source code, customer databases, or crown-jewel stores — using classification labels, path/bucket names, and sampled content. A large transfer of public or low-value data weighs far less than a small transfer of secrets.

3. Destination and channel — signal C

Attribute where it went and how. Signal C fires when the destination is external and unsanctioned — a personal cloud account, a newly-seen domain, an anonymizer, an attacker-controlled host — or the channel is an evasion (archived/encrypted before send, DNS/ICMP tunneling, a covert upload). An internal or sanctioned destination (approved backup, corporate SaaS) weakens it.

4. Actor and intent — signal D

Attribute the transfer to a principal and situate it. Signal D fires when the actor is compromised or acting against pattern: the principal was flagged earlier in an intrusion, a service account moved data it never touches, or a departing employee bulk-pulled right before leaving. This separates theft from a legitimate-but-noisy job.

Verdict rule

  • malicious: signal C plus A or B — sensitive or high-volume data to an unsanctioned destination or through an evasion channel, especially tied to a compromised or against-pattern actor (D). Secrets or a customer database leaving to attacker-controlled infrastructure is malicious on its own.
  • suspicious: an anomalous transfer (A or B) whose destination can't be cleared as sanctioned and whose actor intent can't be established, but benign use can't be shown either.
  • inconclusive: the transfer volume, contents, or destination can't be reconstructed because transfer/DLP logs are missing.
  • benign: the transfer resolves to a sanctioned job — a scheduled backup, an approved data-pipeline sync, a user moving their own files to corporate storage — matching the principal's baseline and role, to an internal or approved destination, with no sensitivity or evasion concern.

Output

Verdict

One of malicious / suspicious / inconclusive / benign, with the signals that decided it.

When not benign: revoke the principal's access to the source store and active sessions, block the destination, preserve transfer logs and any staged archives, notify data-owner and privacy/legal when regulated or customer data is confirmed out, and hunt for the same principal/destination across other stores. Scale notification to the sensitivity found in step 2.

Evidence

The measured volume against baseline, the classification of what left, the destination attribution and channel, the actor and their status, and the window and source store involved.

Reasoning

Name the signals and how they stacked — a service account archiving an entire customer database and pushing it to a personal cloud account it has never used, right after that account was flagged, is malicious; the nightly backup job copying its usual volume to the approved backup bucket is benign.