The brain that decides what to move, and when.
Phorvex evicts attackers from Kubernetes and AI inference workloads in seconds, without breaking what is running. Every decision is mathematically proven, not hand-tuned.
Moving-target defense, finally with a brain.
Conventional MTD rotates on a clock or a rule table. It has no idea what a move costs, what is worth protecting, or what attacker knowledge it actually erases. Phorvex replaces the clock with a decision core that values every possible move and picks the best one under a hard safety budget.
Clocks are predictable
A schedule leaks its own pattern. An adaptive attacker simply learns the rhythm and waits it out.
/02Detection misses what it has never seen
A zero-day emits no alert. Reactive tools sit silent while the attacker settles in.
/03Naive rotation destroys state
Deleting a pod to rotate it kills the database, the training job, the work you were protecting.
Measured on a real Kubernetes cluster, n=20, reproducible, controls clean.
A decision loop that thinks like an economist.
Four stages, every tick, deterministically. The same inputs always produce the same decision, and every decision is recorded with its reasoning.
Scan
Phorvex reads your live fleet and derives its physics: reachability, scarcity, slack, exploitability. Measured, never assumed.
Value
It works out what each workload is worth, how exposed it is, and where the attacker sits on the kill-chain.
Select
The sealed, patent-pending core prices every feasible move and selects the optimal set under a hard safety budget.
Move
Moves run through your unmodified policy controller, with a kill-switch, dry-run, and a complete audit record.
Four engines. One deterministic defender.
Every engine feeds the decision core with measured ground truth. Only the adversarial red-team learns, and it is strictly walled off from decisions.
Physics Engine
Learns your environment by scanning it. Reachability, scarcity, slack, and exposure, measured every tick.
/02Context Engine
Knows where the attacker is and when they will move. A deterministic kill-chain model with a predictive horizon.
/03MTD Engine
Treats moving as an economic decision. Prices every move and selects the provably optimal set under a hard budget.
/04RL Red-Team Engine
A learning attacker that hardens the defender. It already found two real weaknesses, and we fixed both.
Both products from day one.
Kubernetes
Proactive defense that protects production instead of breaking it. Phorvex evicts footholds before any alert fires and never at the cost of a stateful workload. Proven on a real cluster with a reproducible benchmark.
Explore the Kubernetes product Design-complete, GPU validation in progressAI / Inference
The same brain, extended to models, KV-caches, LoRA adapters, GPU partitions, and inference routes. This is attacker state that pod-level tools cannot see. Mechanisms are built and CI-tested. Effectiveness figures wait for real GPU runs.
Explore the AI / Inference productThree outcomes, measured on a real cluster.
Head to head against a blind rotation timer and a reactive-only baseline, on zero-day scenarios that emit no alert. Time to evict the attacker, in seconds. Shorter is better.
| Phorvex | Blind timer | Reactive only | |
|---|---|---|---|
| Eviction dwell on a zero-day | ~5.5s | ~15.2s | ~20.3s, never acts |
| Prevention on a zero-day | 100% | partial | 0% |
| Stateful work preserved | 20 / 20 | 0 / 20 | 0 / 20 |
| Learnable by an adaptive attacker | No | Yes | n/a |
| False positives | 0 | 0 | 0 |
Real Kubernetes cluster, n=20, reproducible, controls clean.
Prove by running. Label everything unproven.
Every capability we claim traces to a real run on real infrastructure. Every figure we have not measured on hardware is marked pending. That discipline is the product.
See the decision brain in action.
A technical briefing, the full benchmark methodology, or a design-partner conversation. Your call.