Researchers propose UniGuardian, a training-free detector for Large Language Models (LLMs) that identifies prompt injection, backdoor, and adversarial attacks without knowing the attack type. UniGuardian measures how prompt perturbations shift the model's output distribution and uses a single-forward strategy for efficient detection and text generation.

CounterSteer is a defense against indirect prompt injection in LLMs, suppressing the behavior by subtracting a learned direction from tool-result tokens during prefill. It requires no fine-tuning, auxiliary models, or added tokens, and achieves 93-100% typography-normalized benign utility while reducing attack success rates to 0.00-0.17.

Researchers identified a vulnerability in large language models to prompt injection attacks, where adversarial content can hijack the model's behavior. They propose a defense by rendering untrusted payloads as images before they reach the model, reducing attack success rates while preserving benign utility.

Researchers found a way to strengthen prompt injection attacks against LLM agents by wrapping injected instructions in the model's own chat template. This allows attackers to evade tokenization-based defenses. The study measured the effectiveness of this technique on various LLM models and found significant improvements in attack success rates.

This paper introduces a new attack method called adaptive long-context prompt injection (AdaLCPI) that reconstructs malicious objectives from incomplete fragments, which can be used to compromise AI agents. This highlights the need for robust safety evaluations of agents against such attacks.