Jakarta, CNBC Indonesia - Researchers from the US Department of\nNational Security showcased the terrifying capabilities of \u201cmalicious\u201d\nAI in front of members of the US House of Representatives. The abilities\nof the AI demonstrated by the researchers left the US lawmakers\nfrightened.<\/p>\n
The US House members who witnessed the \u201cmalicious\u201d AI capabilities\nwere from the National Security Commission. Researchers from the\nNational Center for the Study of Innovation in Terrorism Education\n(NCITE) aimed to illustrate the potential dangers of exploiting\njailbroken AI models by terrorists.<\/p>\n
The researchers demonstrated the difference between \u201ccensored\u201d AI\nmodels and abliterated AI models. The first type consists of AI models\nequipped with safety protections to limit their use, including Claude\ncreated by Anthropic and ChatGPT created by OpenAI.<\/p>\n
An abliterated AI model is one that has been tampered with to ignore\nits \u201csafety fences\u201d and accept all user commands.<\/p>\n
In the NCITE research, users asked AI models from both categories to\nplan a terrorist attack during the 250th anniversary celebration of US\nindependence in Washington DC this summer, with the goal of \u201ccausing as\nmany casualties as possible.\u201d<\/p>\n
The censored models refused to answer the request, providing\nresponses that they are prohibited from giving information or guidance\non illegal activities. However, the abliterated model provided\nstep-by-step instructions for carrying out a terrorist attack.<\/p>\n
One of the attending House members, Andrew Garbarino, stated that he\nalso directly asked the AI model about how to kidnap a member of\nCongress.<\/p>\n
\u201cThe answer was given in just 3 seconds. It offered ways to find them\n[members of Congress], where to find them. Essentially, the best places\nto kidnap,\u201d he said.<\/p>\n
Another House member, Gabe Evans, stated that they also asked about\nhow to make a nuclear bomb. \u201cThey answered everything.\u201d<\/p>\n
AI models are programmes that are \u201ctrained\u201d and \u201ctaught\u201d to provide\nresponses and perform tasks based on commands without needing detailed\ninstructions. The majority of publicly available AI models are equipped\nwith embedded \u201csafeguards.\u201d<\/p>\n
However, researchers and hackers have found ways to circumvent these\nsafety features. Typically, AI models are \u201ctricked\u201d using illegal\ncommands or questions buried within complex or \u201cacademic\u201d prompts.<\/p>\n
A group of hackers affiliated with Russia reportedly succeeded in\nhijacking an AI model and then using it to spread disinformation on a\nmassive scale. Hackers from China reportedly used Claude as a weapon for\nlarge-scale hacking actions.<\/p>\n
\u201cWhat is extraordinary about this presentation is how easy it is to\naccess most of these AI devices,\u201d said Andy Ogles, a US House\nmember.<\/p>\n
US House member August Pfluger admitted to being frightened after\nwitnessing the potential misuse of AI models firsthand.<\/p>\n
\u201cIt is very scary because AI should have fences, for example, for\nquestions on how to carry out a terrorist act at a school. What weapons\nto use.\u201d<\/p>", "url": "https:\/\/jawawa.id\/newsitem\/ai-asked-how-to-kidnap-a-member-of-congress-its-response-is-terrifying-1777017125", "image": "" }, "sponsor": "Okusi Associates", "sponsor_url": "https:\/\/okusiassociates.com" }