One deterministic loop, end to end.
Discover your live fleet. Assess value and kill-chain position. Decide with a sealed, provably optimal core. Act through your unmodified policy controller. No LLM in the decision path, ever.
Scan
Phorvex scans the live environment and derives per-organization physics: the measured properties of your fleet, never hardcoded assumptions.
Value
It computes what each workload is worth, how exposed it is, and where it sits on the attack kill-chain, then projects how long before the attacker advances.
Select
A deterministic core values every feasible move and selects the optimal set under a hard safety and availability budget. The mathematics is intentionally not published.
Move
Chosen moves run safely through your unmodified policy controller or a native actuator, with a kill-switch, dry-run, rollback where possible, and a full audit record.
What makes it trustworthy.
Deterministic
No model in the loop. The decision path is reproducible and explainable. The same inputs always produce the same decision, and every decision is recorded with its rationale.
Safe by construction
Infeasible or SLO-breaching moves are excluded before selection ever happens. Stateful workloads are never destroyed. The feasibility model quarantines them instead.
Adversarially hardened
A learning red-team attacks the deterministic core around the clock. Weaknesses get closed and re-proven, and the learning component never touches decisions.
Multi-tenant safe
One tenant's activity has provably zero effect on another tenant's decisions. Verified under a real adversarial attack, not merely bounded.
You stay in control. Always.
Phorvex is built for operators who get asked hard questions in postmortems. Every control surface is explicit, scoped, and logged.
- Global kill-switch. Halt all actuation instantly, fleet-wide.
- Per-namespace dry-run. Watch what Phorvex would do before it does anything.
- Immutable audit trail. Every decision, taken or not, written to a size-bounded record.
- Least-privilege actuation. An unmodified policy controller via its own APIs, or a scoped per-namespace service account.
Physics Engine
Measures reachability, scarcity, slack, and exposure every tick.
/02Context Engine
A deterministic kill-chain model with a predictive horizon.
/03MTD Engine
Prices every move. Selects the provably optimal set.
/04RL Red-Team Engine
A learning attacker, walled off from decisions.
See the architecture in depth.
The full architecture walkthrough and the technical PRD are available under NDA.