Indirect Prompt Injection: The Foundational Paper on Remotely Controlled AI Assistants
Factos validados
Seminal work by Greshake et al. (Feb. 2023, arXiv:2302.12173). Shows how malicious control instructions can be hidden in web pages, PDFs, or emails to remotely control AI assistants when they summarize those documents.
Explicação simples
A cuckoo's egg in the mailbox: The AI reads your mail aloud, finds a hidden note in the text, and secretly transfers your pocket money to strangers.
Relevância de segurança e dica prática
The fundamental architectural weakness of LLMs: text data and control instructions share the same channel.