When a handles a request, it creates temporary internal numbers as information moves through its layers. These hidden states, including the residual stream (the running representation passed along and updated by each layer), disappear quickly. Ordinary logs and saved model files do not preserve what happened for one request.
The team attached a non-modifying debugger, a tool that observes a running program without changing it, to a graphics processing unit (GPU) process. By tracking memory allocations or GPU work, it located and copied these states without knowing the model’s weights or architecture in advance.
The recovered states helped distinguish refusals from compliant answers, with an area under the receiver operating characteristic curve (AUROC), a measure of how well a classifier separates two groups, of 0.88. But it falsely flagged 28 percent of cases, so the method can flag requests for review but cannot settle them. It worked on tested platforms, not necessarily arbitrary systems.
