Huntra · Detect

Detection that reads every event

Huntra analyses raw, full-fidelity telemetry on GPUs instead of the slice your rules pre-filter. It learns what normal looks like in your environment and flags what doesn't fit.

Pipeline statusExample

tenant acme-eu · Huntra

Ingest rate
4.2 GB/s
Events per second
7.1 M
Entities baselined
48,310
Anomalies today
212
Sent to Veridex
9
Example. Values are illustrative.

Capabilities

What Huntra does

Ingest

Line-rate telemetry

Network flows and packets, identity events, cloud audit logs and endpoint telemetry, read raw.

Baseline

Per-tenant behaviour models

Normal is learned per entity, peer group, time of day and volume, for your environment only.

Classify

ATT&CK mapping

Anomalies are mapped to MITRE ATT&CK techniques with calibrated severity and confidence.

Fuse

One score across sources

Packets, logs and identity signals are scored together, so weak signals add up.

Stream

Sub-second scoring

Design target: anomalies scored in under a second at sustained high throughput.

Learn

Feedback from analysts

Confirmed and dismissed incidents retrain baselines and enrich the shared corpus.

How it works

Inside Huntra

  1. Capture

    Collectors, cloud connectors or a BlueField DPU stream raw events.

  2. Normalise

    Morpheus pipelines parse and enrich events on GPU.

  3. Score

    Baselines and classifiers score each event and entity.

  4. Hand off

    Scored anomalies go to Veridex with their context.

What it reads

Telemetry sources

Network flowsNetFlow, IPFIX, VPC flow logs
PacketsZeek, PCAP, DPU capture
IdentityEntra ID, Okta, Active Directory
CloudAWS, Azure, Google Cloud audit
EndpointEDR process and file events
SaaSMicrosoft 365, Google Workspace
DNSResolver and proxy logs
EmailGateway and mailbox audit

Specification

Inputs, outputs and hardware

ItemDetail
InputsNetFlow / IPFIX, Zeek, PCAP, cloud audit logs (CloudTrail, Azure Activity, GCP Audit), IdP sign-ins, EDR telemetry
OutputScored anomalies with entity context and ATT&CK technique labels
ModelsPer-tenant unsupervised baselines plus supervised and few-shot technique classifiers
HardwareH100 / H200 for training, A100 / L40S for streaming inference, optional BlueField DPU capture
SoftwareNVIDIA Morpheus, RAPIDS (cuDF, cuML), Triton, TensorRT

Questions

About Huntra

How is this different from SIEM correlation rules?

Rules match patterns someone wrote in advance. Huntra learns normal behaviour for your environment and flags departures from it, including techniques no rule covers yet.

Won't anomaly detection be noisy?

Anomalies alone are noisy, which is why Huntra fuses sources and hands its output to Veridex to investigate before anything reaches an analyst.

Do baselines from other customers affect ours?

No. Baselines are per tenant. Only anonymised technique patterns are shared, and only if you opt in.

Point Huntra at your own telemetry

We'll scope an assessment with you: which sources {name} would read, for how long, and what we'd report back.