A learning attacker that hardens a deterministic defender.
The only machine learning in Phorvex is an adversarial reinforcement-learning red-team that stress-tests the decision core around the clock. It is strictly walled off from decisions and human-gated.
Continuous co-training
The red-team trains against the live defender, probing for structural weaknesses a human team would take months to find.
Walled off from decisions
Learning never touches the decision path. The defender stays deterministic, reproducible, and explainable. The red-team only attacks it.
Human-gated
Findings become fixes only through human review. Each weakness is closed as an auditable change and re-verified to neutralization.
It has already earned its keep.
The red-team has found two real structural weaknesses in the decision core, and we fixed both. Each was closed as an auditable, regression-gated change and re-verified to neutralization.
That is the loop working as designed. A learning attacker makes the deterministic defender stronger, without ever becoming part of it.
Why no ML in the defender?
A learned defender can be probed, drifted, and fooled, and it cannot explain itself in a postmortem. We put the learning where it belongs: on the attacking side, where its only job is to break our assumptions before a real adversary does.
Ask us what the red-team found.
The two structural findings, and how each was closed, are covered in the technical briefing.