Skip to main content

Data exfiltration detection

How an AI SOC detects exfiltration to personal cloud storage: baseline each user's uploads, flag spikes, add HR context, and tune out approved storage.

Use case
Updated

The short answer

Exfiltration detection looks for data leaving through channels that blend in, such as uploads to personal cloud storage. Cotool's example detection models each user's normal upload volume from proxy logs, flags days far above it, and joins the spikes with HR data on who is leaving.[1][4]

The problem

Exfiltration to services like Dropbox or Google Docs gives an adversary cover when hosts already talk to those services.[4]

One fixed threshold doesn't fit everyone, because normal upload volume differs from user to user. The example detection models each user's own baseline instead.[1]

How an AI SOC handles it

  1. Query proxy logs

    The detection reads 30 days of uploads to personal storage from web proxy logs, such as Zscaler logs sent by a Cloud NSS feed.[1][5]

  2. Model each user's normal

    It computes a 30-day rolling median per user and flags days more than 20 times that user's normal.[1]

  3. Flag leavers

    Spikes are joined with each user's last day from the HR system.[1]

  4. Tune out approved storage

    In version 2, the tuning agent added an exclusion for the company's own Box storage.[1]

  5. Investigate each alert

    Each remaining spike becomes one alert. The agent checks the user's Google Workspace Drive activity and recent sign-ins before deciding.[1][2]

Example

Exfiltration to personal cloud storage (T1567.002)text
LOOKBACK = 30   # days of history per user
SPIKE = 20      # flag days this many times their normal
CATEGORY = "personal-storage"

1. Query proxy logs     uploads to personal storage, 30 days
2. Model each user      30-day rolling median per user
3. Find spikes          days over 20x the user's normal
4. Flag leavers         join each spike with a last day
5. Exclude company Box  added by the tuning agent (v2)
6. Return alerts        one alert per remaining spike

The parameters and visual workflow of the example detection on Cotool's Detect page. Adjust the lookback, spike factor, and storage category to your traffic.[1]

What people approve

  • Exclusions a tuning agent proposes are sent for review or applied by a person.[1]
  • Response agents cannot close an alert as True Positive. When an agent confirms malicious activity, it escalates the alert and a person closes it.[3]

How Cotool handles it

  • Cotool detections are Python that runs against stored and federated sources, and each one maps to a visual workflow the whole team can read.[1]

Integrations used

Sources

  1. Detection engineering (Cotool Detect) · accessed November 16, 2026
  2. Changelog (Cotool docs) · accessed November 16, 2026
  3. Alerts (Cotool docs) · accessed November 16, 2026
  4. Exfiltration to Cloud Storage, T1567.002 (MITRE ATT&CK) · accessed November 16, 2026
  5. About Cloud NSS feeds (Zscaler Help) · accessed November 16, 2026

Related use cases

All use cases

Related reading

Attackers are scaling with tokens

Cotool helps defenders operate at machine speed. See how security teams are scaling Detection & Response.

Book a demo