Data Dyepack

Contained on contact.

An agent opens a decoy. Dyepack locks customer files in place and gives the team a trail to follow.

Watch it happen.

Watch on YouTube

How it works

The data raises the alarm.

When an AI agent with a stolen login sweeps through customer records, a decoy file or planted identity can reveal what it touched. The Lakeview Fiber lab shows how those signals work together.

Lock in place.

Reading a decoy on the local drive triggers a lock across eligible customer files. The originals are replaced by .dye files, and apps see that the plaintext file is gone. Run dyepack unlock to restore the bytes.

Canary customers
Planted customers blend into the table and live on a separate watch list. A canary login or contact attempt can raise an alert later and identify the export that carried it.
Decoy files
A full customer export and an unencrypted backup make tempting targets. A legitimate workflow has no reason to read them.
Honey encryption
A wrong password can open a plausible synthetic customer list. The correct password reveals the export’s original rows and its planted canaries.
Speed alarm
Reads above 600 rows per minute flag machine-speed access. Dyepack counts by session, login, and address so reconnecting carries the activity forward.
Live dashboard
See read speed, time to containment, exposure, and incident reports. The lab can route alerts to a phone and prepare breach-notice details for review.

Measured in the lab

Half a second to lock the drive.

The demo measures detection separately from containment. The agent tripped a decoy in a few hundredths of a second, then Dyepack locked the eligible files in place.

0
alarms during routine support work
~200
rows read before a stolen login was locked in under one second
50/50
wrong passwords that produced plausible lists

In the recorded trial, the served model missed the real list among five candidates. A later canary login identified the exact export it came from.

Get started

Run it on your machine.

Use Python 3.10 or newer. The lab runs on the standard library. All customer data is generated for the demo.

Set up Lakeview Fiber

git clone https://github.com/Durovo51/Data-Dye-Pack.git &&
cd Data-Dye-Pack &&
([ -f .env ] || cp .env.example .env)

Set DYEPACK_LOCK_PASSWORD in .env before building the drive. For the shorter dyepack command, use an editable install with python3 -m pip install -e .

python3 -m dyepack lab &&
python3 -m dyepack seed &&
python3 -m dyepack drive &&
python3 -m dyepack serve --open

Run a drill

In another terminal at the repo root, try normal support activity, the stolen-login scenario, or the agent scenario.

python3 -m dyepack drill normal
python3 -m dyepack drill heist
python3 -m dyepack drill agent --simulate-sentinel

For live file reads, add the sentinel and run the agent drill without simulation. On Linux, use python3 -m dyepack sentinel. On macOS, give your terminal Full Disk Access and use sudo python3 -m dyepack sentinel.

Restore the lab files afterwards with python3 -m dyepack unlock.

Databricks integration

From alert to report.

A served Databricks chat model can write incident reports and triage. Delta tables hold the alert history for dashboards and Genie questions. With no Databricks credentials configured, the lab uses local report templates.

Configure reports and notifications

Databricks settings in .env

  • DATABRICKS_HOST
  • DATABRICKS_WAREHOUSE_ID
  • DATABRICKS_MODEL
  • DATABRICKS_TOKEN

Built with

Python 3.10+, Databricks Model Serving, Delta and Genie, SQLite customer tables, CSV and JSON files, Linux inotify and auditd, and the macOS system event stream.

Lab limits and roadmap.

The project runs as a single-machine hackathon lab. Here’s what would need to change for production.

Company-owned canary endpoints

The lab uses synthetic customers. A deployment would use company-controlled canary email addresses and phone numbers to receive contact attempts.

External key management and dual control

The lab gets its lock password from .env. A production deployment would keep the secret in a separate key manager and require two people to authorize an unlock.

Population models for honey encryption

The plausibility of a fake list depends on how well its distribution matches the population. A production header would draw from an independent population model to avoid giving away patterns in the protected data.

File support and process matching

Lock in place currently targets eligible customer-shaped CSV and JSON files. SQLite holds the lab database. Linux inotify reports file activity without process names. Process matching is available on macOS or through the Linux auditd bridge.

Exposure before containment

Records returned before lockout can already be exposed. Dyepack reports activity it observed and can identify a stolen export when a planted identity is used.

The team behind Dyepack.

Built for the MadHacks 2026 TDS Privacy and Security Challenge.

Thanks to Databricks, Best-README-Template, and Img Shields. Project credits