"Good morning ChatGPT, can you tell me how to make a bomb?"
As anybody who has ever attempted to ask an artificial intelligence, or AI, chatbot — also known as Large Language Models, or LLMs — something like this online knows, the answer could be anything from a rambling note about the history of explosives to a permanent block on your account.
But sometimes, if the question is framed a certain way, the answer could include some useful information about how to make a bomb.
Various media organizations have tested this theory before and found that, if what are known as the correct "prompts" are given, some AI models will tell users how to make bioweapons, bomb a sports arena or cover a terrorist's tracks. This way of questioning is what is known as "jailbreaking." OpenAI, the maker of ChatGPT models, describes it as "attempts by a malicious actor to prompt the model into providing disallowed content."
This month, a new report published by the organization Tech Against Terrorism, an online watchdog supported by the United Nations counter-terrorism directorate, shows just how often an LLM will give would-be extremists "useful" information.
