Dolaylı Prompt Injection Attacks are now defined as Test Time Search Problem

The threats for the safety of artificial intelligence models are becoming increasingly sophisticated. A new research published in arXiv is formulated in a completely new way of indirect prompt injections, and it defines attacks as a search problem in the test time.

Basic findings of research work:

-Amidable attack surface: Attacks are no longer only directly on the surface of an attack, connected to the environmental, user task and injection task.

• Agentic aggressive architecture: The researchers developed a aggressive system that makes environmental discovery, performs structured thinking on attack strategies and can be adapted to the response of the victim agent.

• Real-time adaptation: The system combines dynamically attack methods by interacting with the victim agent.

• Tested in various tasks: Research is tested on heterogenic tasks, providing general validity.

This formulation allows artificial intelligence security researchers to systematically understand and meet prompt injection attacks. Agentic attacker approach makes it possible to test more robust than defense mechanisms by increasing the complexity and reality of attacks.

The findings light on how safety losses can occur especially in multilayer AI systems. Thus, it can contribute to the development of more resistant AI models in the future.