How does Vardar connect without adding inline control risk?

Vardar begins with a read-only copy of traffic from SPAN or TAP and can add customer-approved read-only evidence sources. Active integrations remain explicitly governed by customer policy and approval.

Does raw OT data leave the site?

The Vardar Edge AI appliance performs decoding, correlation, inference and investigation locally. Raw OT traffic and process values remain inside the customer environment.

How does Vardar reduce false positives?

Vardar compares two evidence paths: network behavior and process or physics behavior. Agreement raises confidence; a single evidence path stays under investigation instead of being presented as confirmed cyber.

Attack #10, 1P005ON + 1P006OFF (WaDi 2017)

1.0Plant

Where it hit the plant

primary grid, raw-water intake & dosing

1P005ON1P006OFF

elevated reservoir, secondary distribution

consumer / return network

2.0Timeline

Attack window and detections

Labelled window2017-10-11 10:55:00 to 2017-10-11 10:56:27plant local time

Physical Residual

Shaded band: the labelled attack window from the iTrust ground-truth catalog. Marks: the first second each engine flagged the window. "No flag" means the engine stayed silent for the whole window.

3.0Evidence

Why it was flagged

At 2017-10-11 10:55:00, attack #10 (SSMP) manipulated 1P005ON, 1P006OFF — “Pipe bursts”. The serving model raised the alert 7 seconds after onset (peak score 3.4178x the operating threshold). The dominant physical evidence: 1P006 — raw-water booster pump P6, Stage 1 (primary grid — raw-water intake & dosing) was driven to a state the plant's learned physics did not predict from the rest of the process, carrying 79% of the anomaly score — 1226x its share in normal operation. Supporting deviations were concentrated in 2FIT001, 1P005, 2MV003. In plain terms: the plant's own physics stopped agreeing with what the sensors were reporting, in exactly the part of the process the attacker touched.
From the official explain run 2026-06-14T18:37:06Z

Peak score 3.42x of the operating threshold

Sensors carrying the anomaly score

Sensor	Share of score	Deviation	Share vs normal
1P006 raw-water booster pump P6, Stage 1 (primary grid — raw-water intake & dosing)	79.3%	198.9σ	1226.0x
2FIT001 flow meter, Stage 2 (elevated reservoir — secondary distribution)	2.4%	6.2σ	3.1x
1P005 raw-water booster pump P5, Stage 1 (primary grid — raw-water intake & dosing)	2.0%	5.1σ	2.8x
2MV003 elevated-reservoir feed motorized valve, Stage 2 (elevated reservoir — secondary distribution)	0.8%	2.2σ	1.2x
2AAIT003 plant instrument, Stage 2 (elevated reservoir — secondary distribution)	0.5%	-1.2σ	0.2x
2_FIC_201_SP plant instrument, Stage 2 (elevated reservoir — secondary distribution)	0.5%	-1.1σ	0.2x
3AIT004 water-quality analyzer, Stage 3 (consumer / return network)	0.4%	1.1σ	0.4x
2AAIT004 plant instrument, Stage 2 (elevated reservoir — secondary distribution)	0.4%	-0.9σ	0.2x

4.0Provenance

Every figure has a run.

Validation run

2026-06-13T22:27:18Z, commit de79dac5398150de9bc5eda89a529d2dbb942c6c

Model image

sha-de79dac53981

Metrics artefact SHA-256

a05be50d6a12ad442d292f0b9759d4087430c537fcc3b93712812ff55d5605bb

Explain run

2026-06-14T18:37:06Z

Explain artefact SHA-256

ce5f77ae0bc78a0576ec3ca053b5594844039f7f77c0aa2ba63850a9bf46874f

Generator

vardar-portal/scripts/build-gallery-data.mjs (DEL-34/DEL-43)

SWaT and WaDi datasets courtesy of iTrust, Centre for Research in Cyber Security, Singapore University of Technology and Design. Detection figures trace to the reproducible model runs shown with each dataset.

More caught attacks from WaDi 2017Attack #2 Attack #16 All caught attacks

The next record is your process.

A technical evaluation on one critical process, with success criteria agreed in advance, produces the only figure that applies to your site.

Evaluate one critical process

Attack #101P005ON + 1P006OFF

Sensors carrying the anomaly score

The next record is your process.