On the Geometric Limits of Transformer Defenses against Obfuscation Attacks: Latent Embedding Collapse & Performance Robustness Gap

Abstract

Prompt injection attacks pose significant risks to language model safety, yet existing defenses are typically evaluated using classification performance. We show that high detection performance does not imply representational robustness. Specifically, multi-operator obfuscated prompts (combining homoglyphs, zero-width characters, and punctuation or emoji noise) can partially collapse onto the embedding manifold of clean prompts, a phenomenon we term latent embedding collapse. Results indicate that across multiple BERT family encoders with varying depth and capacity, detectors achieve near-perfect classification performance, yet the minimal clean-obfuscated margin delta = 1.02, indicating near-overlap of obfuscated and clean embeddings. Obfuscated embeddings further exhibit elevated intra-class variance (3.33 +/- 6.23), indicating severe latent-space instability despite high performance. These results reveal a substantial perf ormance-robustness gap, demonstrating that standard evaluation metrics fail to capture latent embedding collapse and underlying geometric fragility. Our findings show that increasing model capacity does not eliminate latent embedding collapse, motivating geometry-aware robustness analysis as a necessary complement to performance-based evaluation for prompt-injection defenses.