Blog

Notes from the
AI security frontier.

Research, threat write-ups, and field notes from assessments — published as we learn them, not on a schedule.

Why AI security has to happen at the response level

Input filtering catches the attacks you expect. It misses the ones that only show up once the model has already answered.

A walkthrough of how a single compromised document can quietly change what your retrieval-augmented model tells every user.

Semantic entropy gives you a number for how confident a model actually is — not just how confident it sounds.

When a model can call tools, book things, and send messages, a bad response isn't just wrong text anymore — it's a side effect.

Regulators are starting to ask for evidence, not intentions. Here's what that shift means for how security teams should prepare.

A look at which jailbreak patterns are still getting through production defenses, and why the simple ones are often the hardest to catch.