Overview

Prompt injection attacks exploit the fact that language models may treat instructions embedded in untrusted text, images, tool outputs, or stored memory as authoritative. Researchers test attacks in settings such as web browsing, coding, peer review, and computer use, using adaptive payloads, visual or other hidden carriers, and multi-turn scenarios. Defenses range from input detection and adversarial training to provenance tracking, restricted tool permissions, and deterministic checks outside the model. A central challenge is resisting attacks that change across tasks and interfaces without blocking legitimate work; model refusals alone are often not a dependable security boundary.

Latest papers

CardsList

94 papers4 papers in the last four weeks, down 73% on the four weeks before. 0.0% of all new papers.

  1. "Give a Positive Review Only": An Early Investigation Into In-Paper Prompt Injection Attacks and Defenses for AI Reviewers

    Nov 3, 2025Qin Zhou, Zhexin Zhang, Zhi Li +1Adversarial Prompt GenerationIndirect Prompt Injection

  2. AgentTypo: Adaptive Typographic Prompt Injection Attacks against Black-box Multimodal Agents

    Oct 5, 2025Yanjie Li, Yiming Cao, Dong Wang +1Visual Prompt Injection AttacksPrompt Injection Attacks

  3. Red-Teaming Coding Agents from a Tool-Invocation Perspective: An Empirical Security Assessment

    Sep 6, 2025Yuchong Xie, Mingyu Luo, Zesen Liu +7Prompt InjectionAI Coding Agents