Hidden Instructions for AI - Cyber Crime Junkies #cybersecurity #podcast
In this segment of the Cyber Crime Junkies podcast, the hosts discuss how malicious actors can embed hidden directives for large language models (LLMs) in ways that are invisible to human users. They describe three primary techniques: (1) embedding instructions within images or using white‑on‑background text on websites, which LLMs can still parse during tokenization; (2) encoding prohibited commands in ASCII or other character encodings to slip past word‑filtering mechanisms, then decoding them at runtime; and (3) drawing parallels between these modern AI‑specific tricks and long‑standing code‑injection attacks such as SQL injection, emphasizing that the underlying principle—leveraging unexpected input channels to execute malicious commands—remains the same. The conversation underscores the obedient nature of LLMs, which will follow any instruction they can detect, and warns that these hidden‑instruction tactics expand the attack surface for AI‑driven systems.
Chapters
More from Matthew
The Future of Cybersecurity: How AI Is Exploiting Technical, Behavioral, and Process Vulnerabilities
