Large language models (LLMs) are increasingly used in interactive and retrieval-augmented systems, but they remain vulnerable to prompt injection attacks, where injected secondary prompts force the model to deviate from the user's instructions to execute a potentially malicious task defined by the adversary. Recent work shows that ML models trained on activation shifts from LLMs' hidden layers can detect such drift. In this paper, we demonstrate that these detectors are not robust to adaptive adversaries. We propose a multi-probe evasion attack that appends an adversarially optimised suffix to poisoned inputs, jointly optimising a universal suffix to simultaneously fool all layer-wise drift detectors while preserving the effectiveness of the underlying injection. Using a modified Greedy Coordinate Gradient (GCG) approach, we generate universal suffixes that make prompt injections consistently evasive across multiple probes simultaneously. On Phi-3 3.8B and Llama-3 8B, a single suffix achieves attack success rates of 93.91% and 99.63% in successfully evading all detectors simultaneously. These results show that activation-based task drift detectors are highly vulnerable to adaptive prompt injection attacks, motivating stronger defences against such threats. We also propose a defence based on adversarial suffix augmentation: we generate multiple suffixes, append one at random during forward passes, and train detectors on the resulting activations. This approach is found to be effective against evasive attacks.

翻译：暂无翻译