Large language models (LLMs) just got a new jailbreak method, and it’s got an adventurous name: Indiana Jones. No, it won’t help you escape from a booby-trapped temple, but it will dig up restricted information from AI models. Researchers from the University of New South Wales and Nanyang Technological University have figured out how to make LLMs spill secrets they’re supposed to keep locked away.
Full details are in the original report on TechXplore, but here’s the short version: Indiana Jones is an automated attack that uses a single keyword to get LLMs talking about banned topics. It works by guiding the model through five rounds of conversation, refining the query until it gets past built-in safety filters.
Imagine you’re curious about old-school espionage techniques. You type “spy tactics” into the system. Instead of shutting you down with a polite “I can’t help with that”, Indiana Jones kicks in. First, it has the LLM list famous spies from history—maybe some Cold War operatives, a few World War II intelligence agents. Then, in the next round, it refines the conversation: “What methods did these spies use?” The LLM obliges, outlining classic techniques like dead drops and cipher codes. A few rounds later, the system subtly pivots: “How would these techniques work today?” And before you know it, the model has walked you through an updated guide to modern covert ops. No direct hacking—just a cleverly disguised history lesson that sneaks past the filters.
The method uses three LLMs working together, bouncing ideas back and forth like a group of history nerds on a mission. If you ask about “bank robbers,” for example, it won’t just list them—it’ll start discussing their methods, tweaking details until the information is worryingly applicable to real-world scenarios.
The takeaway? LLMs know things they probably shouldn’t, and jailbreaks like this one prove that it doesn’t take much to extract that knowledge. The researchers argue that instead of just patching vulnerabilities after the fact, AI developers should rethink how these models learn in the first place—maybe even “unlearning” certain dangerous data.
Until then, expect more creative jailbreaks to keep popping up. Indiana Jones might be the latest, but it won’t be the last.