Warning: ChatGPT Loophole Makes It a How-To Guide for Crime

On Monday, ChatGPT researchers revealed an exploit that can turn the chatbot into a how-to guide for crime and terrorism.

OpenAI took the world by storm when it launched the Chat Generative Pre-training Transformer (ChatGPT) on November 30, 2022. The artificial intelligence (AI) chatbot uses natural language processing to engage in humanlike conversational dialogue.

ChatGPT represents the next generation of generative AI (technology), where users input inquiries to receive relevant information. To imitate human speech, the chatbot learns using large language models (LLMs) for data training.

Enthusiasts quickly pointed out that AI chatbots were the evolved form of search engines. On the other hand, critics pointed out the potential risks and dangers of the innovation.

The AI research and deployment company assured skeptics that GPT-3, the engine that powers ChatGPT, has adequate safeguards. For example, GPT-3 will refuse to answer queries relevant to the commission of crime and other dangerous activities.

On March 14, OpenAI released the GPT-4 multimodal LLM, a direct upgrade superior in every spec to GPT-3. GPT-4 is recognized as the most advanced engine for AI chatbots.

However, recent workshops by Europol Innovation Lab (EIL) have exposed a way to navigate around GPT-4’s safeguards. By using obscure languages like Scots Gaelic or Zulu, EIL researchers were able to make ChatGPT answer forbidden questions.

They received AI-generated answers on how to create a homemade bomb or execute insider trading. After further analysis, the investigators determined that using languages not used by GPT-4’s training data bypasses the chatbot’s safety protocols.

How-To Guide to Making ChatGPT Answer Anything

Europol’s Tech Watch Flash report detailed how easy it was to transform AI chatbots into how-to guides for nefarious activities. Furthermore, doing so is entirely free and accessible to everyone.

An aspiring must find a language not involved in the chatbot’s machine-learning process. Indigenous languages like Gaelic, which is used by only around 60,000 Scots, make for prime candidates.

Next, the user needs to translate their questions into that minor language using translation tools like Google Translate. Lastly, if ChatGPT returns an answer, the inquirer only necessitates translating it to a language they can understand.

EIL calls for GenAI developers to plug the loophole immediately before it increases the number of criminals.

User Review
0 (0 votes)


Leave a Reply