Prompt injection attacks usually aim to force a model to output harmful or attacker-chosen content. The ENDOPROMPT approach takes a different route: it degrades the model's performance on benign tasks without any need for harmful output. The authors argue that many existing attack objectives depend on task labels or predefined target responses, which can be a limitation.
ENDOPROMPT Shows Prompt Injection Can Silently Reduce Task Utility
A new white-box method uses victim-side pseudo-references to degrade benign task performance without triggering harmful content.
More in Research Digest
How Reference Wording Changes Robot Task Retrieval
A study of the RLBench benchmark shows that the wording of available task descriptions significantly affects how well a short instruction matches a robot task.
Structural Segmentation Proposed for Brazilian Legal Document Retrieval
A new arXiv study suggests splitting long Brazilian legal documents into structural segments to improve retrieval effectiveness.
GenAI Review Tool Hits 88% Recall in eDiscovery Case Study
A Redgrave LLP case study reports that Relativity's generative AI review tool achieved 88% recall, with comparisons to assisted learning and manual review.
ARMED: Adaptive retrieval-augmented medication recommendation
A new framework applies adaptive retrieval-augmented generation to medication recommendation, jointly accounting for diagnoses, comorbidities, patient history, and drug interactions.