New Framework Diagnoses Classifier Decisions for Prompt Injection

Developing reliable classifiers, especially for critical tasks like detecting prompt injection in large language models, requires not only high accuracy but also an understanding of *why* a classifier makes certain decisions and when those decisions can be trusted. Traditional methods often lack this diagnostic capability.This research proposes the Latent Diagnostic Taxonomy, a comprehensive framework for both constructing optimized classifiers and developing a complementary diagnostic tool. The framework involves three key steps: first, building a dimensionality-optimized classifier where embedding dimensions are empirically selected; second, identifying a small set of "latent support vectors" that represent influential prompts for altering predicted labels; and third, using these influential tokens and their attack magnitudes to construct a diagnostic taxonomy.Applying this framework to a prompt injection classifier, the study found that a significant portion (77%) of confident decisions were brittle, meaning they could be flipped by removing a single token. This brittleness revealed two distinct failure patterns: confidence calibration issues and genuinely exploitable shortcuts. The diagnostic taxonomy categorizes prompts for different treatments: "Safely" rely on the classifier, flag "Heuristic Bias" or "Heuristic Override" cases, or route "Insufficient Context" cases for human review, providing clear remediation strategies for each category.