Prompt Injection Attack
A prompt injection attack is a security attack targeting large language models (LLMs) and their applications. Attackers embed carefully crafted malicious instructions within input content to interfere with or override the model's original system prompt, manipulating the model into executing unintended behaviors such as bypassing safety restrictions, leaking sensitive information, or performing harmful operations. This attack exploits the inherent limitation of models in distinguishing normal user input from malicious instructions.
Core Facts
Timeline (last 90 days)
Prompt injection attacks do not directly attack the bank system; instead, they use the AI agent's proxy identity as a "legitimate user" to bypass the bank's risk controls.
All Facts (2)
Prompt injection attacks fundamentally exploit the inherent flaw of large models in failing to separate instructions from data, and are divided into two categories: direct injection and indirect injection.
70%UnverifiedPrompt injection attacks do not directly attack the bank system; instead, they use the AI agent's proxy identity as a "legitimate user" to bypass the bank's risk controls.
50%