Stabilize AI Geometry, Shrink the Attack Surface

I’ve been running a series of experiments exploring what happens when you stabilize the internal state space of AI models. The idea is fairly simple: if the model’s geometry is stabilized during training, or its state space is constrained from drifting during inference, its behavior remains inside a tighter, more predictable envelope.

The results have been interesting.

In controlled comparisons with identical architectures, the stabilized variant accumulated noticeably less KL-divergence during generation, produced far fewer curvature spikes in the deeper layers of the network, and held together under chained perturbations that normally push a standard model into repetitive collapse.

What surprised me wasn’t just the stability itself, but the number of problems that seemed to shrink alongside it. Hallucination cascades became less likely. Adversarial prompts had less leverage to push the model into alternate trajectories. Even the familiar failure mode where a system gradually derails under recursive reasoning pressure showed up later and with less severity.

It begins to suggest that many of the issues we treat as separate—hallucinations, jailbreaks, prompt injection, recursive derailment—may not be independent problems at all. They may simply be different symptoms of the same underlying instability in the model’s internal geometry. If that’s true, then stabilizing that geometry doesn’t just make the model behave “better.” It reduces the space in which those failures can occur in the first place. Which raises an interesting possibility: some of the problems we’re trying to solve with guardrails and policy layers might actually be architectural problems hiding underneath them.