Opening article
The AISI Blog has published an introduction to its work on white box control along with an update on its current research. The post provides foundational context and outlines the organization's progress in this specific research area. Readers are given a direct look into the ongoing efforts surrounding white box control.
AI control aims to design measures that will prevent future systems from causing harm, under the worst-case assumption that they are not aligned with our interests.
We can sort control techniques into two categories: black box measures and white box measures. Black box control relies on monitoring a model’s external behaviour, such as its chain-of-thought and final outputs. White box control, on the other hand, involves observing or modifying its internal activations.
At the AI Security Institute (AISI), we’ve assembled a dedicated White Box Control Team. Our current focus is on detecting sandbagging , a phenomenon where models strategically underperform during evaluations. Today, we’re releasing a detailed progress update describing some preliminary results.
In this blog, we introduce the concept of white box control and provide some brief insights from our research so far.
White box control involves altering or observing what’s going on inside a neural network in an effort to prevent harm from misaligned models.
It is an application of mechanistic interpretability , a field which aims to understand the inner workings of AI models. White box control and mechanistic interpretability can employ many of the same tools, such as linear probes (which can isolate representations within specific layers of a network) and sparse autoencoders (which help disentangle neural networks into a form that is more interpretable).
Understanding white box control research from organizations like AISI provides insight into evolving technical methodologies and transparency efforts within the artificial intelligence sector.