5.5 seconds, and nothing deleted

A zero-day does not announce itself. So we built a benchmark around the one situation where most defenses have nothing to react to: an attacker who is already inside, and an alert that never comes.

The setup

The runs happened on a real Kubernetes cluster, not a simulator. Each scenario plants a foothold that emits no signature and trips no alert. We repeated it twenty times and kept the controls clean, so the result is reproducible rather than a lucky run.

Three defenders faced the same scenarios:

• Phorvex, which decides what to move and then moves it.
• A blind rotation timer, which moves pods on a fixed schedule.
• A reactive-only tool, which waits for an alert before doing anything.

The results

Time to eviction is the headline, but it is not the number we care about most. Two others matter as much:

• Prevention. Phorvex prevented the attack in 100% of the zero-day scenarios. The timer managed a partial result. The reactive tool scored 0%, because an alert that never fires never triggers anything.
• Stateful work preserved. Twenty of twenty stateful jobs were still running after Phorvex acted. Both baselines preserved zero of twenty.

Evicting an attacker by deleting the workload they sit in is not a win. It is the attack finishing the job for them.

Why "nothing deleted" matters

A timer does not know what a pod is doing. When its turn comes, it deletes it, whether that pod is a stateless web server or a training job that has been running for hours. The foothold goes away, and so does the work.

Phorvex refuses any move that would destroy state. When the foothold is inside a stateful workload, it quarantines that workload instead: the attacker is cut off and the job keeps running. That is where the 20 / 20 comes from.

False positives

All three defenders recorded zero false positives in these runs. For the reactive tool that is trivially true, since it never acted. For Phorvex it matters because it acts without waiting for an alert, so it has to act only when it should. The benchmark also found that its moves are not learnable by an adaptive attacker, unlike a schedule.

What this does not claim

If you want to go through the method in detail, including the scenario design and the controls, that is exactly what a technical briefing is for.