In the AI security landscape, prompt injection attacks have rapidly become the most significant threat. Large language models cannot fully distinguish legitimate user instructions from malicious commands hidden in emails, code, and various content; this enables exploitation and the execution of hackers' commands.
Since there is no reliable boundary between trusted and untrusted sources, AI engine developers are forced to design only temporary protective measures instead of fundamentally solving the problem. So far, most prompt injection attacks have been of the "push" type; that is, hackers must send the malicious command to each target individually, which limits the scale of the attack. However, if these limitations are overcome, the way will be paved for widespread exploitation of this vulnerability and the creation of botnets.