Are You Reading the Right Thing? Rethinking Internal Signals for Prompt-Injection Detection
The paper compares five ways of reading LLM internals for indirect prompt-injection detection, showing that probes, attention-based measures, role/exposure signals, and concept readouts are not all measuring the same security-relevant property. Its useful twist is separating “does an injection exist?” from “does it actually work?”, finding that signals closer to response formation—especially attention diversion—track successful attacks better, though the evidence is still limited to a structured setup, two models, and results that can shift a lot with the prompt template.
Abstract
Several recent studies propose detecting prompt injections from a language model's internal state. However, these methods do not measure a single security property. Residual-stream probes, attention scores, role projections, exposure probes, and concept readouts may capture injection presence, recognition, information routing, or behavioral influence. It is therefore unclear what differences in prompt-injection detection performance reveal about the internal signals these methods read. We study how five representation-reading methods vary under a common structured indirect prompt-injection setup on Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it. We evaluate each method by how well it distinguishes injected prompts from benign controls. We then further analyze each detection method's ability to detect attacks in cases where they are successful and in cases where they are not. This helps us better assess what properties each detection method might rely on. Methods trained to detect injection presence or attack-related concepts generally separate unsuccessful injections more strongly from benign inputs, whereas attention diversion, exposure, and role-confusion-based methods separate successful injections more strongly on both models. Attention Tracker provides the strongest consistent successful injection separation, reaching 0.949 AUROC on Qwen and 0.887 on Gemma. These results suggest that signals closer to response formation may better encode information about an attack's success, and that detection performance can depend strongly on the prompt template. We share the data associated with this study at https://anonymous.4open.science/r/paper-evaluation-data-7877