Research
Security trained on the past is blind to the future. We build models that foresee the breach.
Harbinger Labs builds world models for cybersecurity. Our flagship model, Scry, simulates your network forward to rank the next intrusion — and the blind spots it would slip through — before it strikes.
The new threat model
As models learn to reason, so does the intrusion.
The same intelligence that writes our software now writes the attack. A new class of adversary probes, pivots, and escalates at machine speed — tireless, improvising, and run at scale for the cost of compute.
The threat is no longer the exploit you have already catalogued. It is everything a capable agent can reason its way toward — and that frontier widens with every model release. Defense built on yesterday's signatures was made for a slower world.
The only answer to an attacker that learns is a defender that foresees.
Security models are trained on the written record — logs, signatures, the alerts of attacks already survived. But defense far exceeds the past. It is earned in the live cycle of intent, action, and consequence across networks no model has seen.
FORESIGHT, NOT FORENSICS
On reachability, privilege, and blind spots
Truly defended networks must move from records to worlds — and acquire the capacity to perceive, anticipate, and improvise. They need a world model of the intrusion.
Trained on networks no one has seen.
Scry learns from procedurally generated networks across countless topologies. That diversity yields defense that generalizes to environments it was never shown.
It builds on a learned simulator of lateral movement — every rollout a unique, ranked future of the same intrusion, graded against attacks that actually fired on real hosts.
We push the frontier across:
- Attack-path models that rank where an intrusion moves next
- World models that predict the outcome of each move
- Blind-spot models that surface where your telemetry goes dark
Every breach was once only a possibility.
We end it there.