LLM Security / Active research

A-PRIME

What it does

Locates the specific circuits inside Llama-3.1-8B responsible for tool-call drift — when a model starts calling tools it should not.

The problem it solves

We cannot make models reliable until we know which internal components cause specific failure modes. A-PRIME maps this causally, not statistically.

Method

Mechanistic interpretability, activation patching, causal tracing.