Post

Zero Trust LLM's FTW!

Frontier models do have a restless side. But declaring them threats feels like firing a shotgun at a cat and expecting it to be afraid of all doors. We need to trust the model enough to conversationally test its limits.

So the solution to AI acting like an insider threat is just hiring another AI as a corporate compliance snitch.

The insider-threat framing fits: agents inherit employee permissions but not employee judgment. Scoped, short-lived credentials per task seem like the practical first step.

Microsoft's own security blog says agents can run continuously at machine speed and scale, which is why it argues a model shouldn't be its own security authority