- By Vikas Yadav
- Sat, 29 Jul 2023 03:14 PM (IST)
- Source:JND
ChatGPT, Bard and other AI models have made headlines all over the internet since their launch because of their AI capabilities. While the tech moguls harnessing them have stressed their commitments towards safety and ethical use, researchers have discovered a method that could potentially trick these models to bypass these guardrails and perform harmful actions without adhering to these guidelines.
Top players in the AI space such as Google, Anthropic, OpenAI and more, have advocated AI safety considering the concerns posed by tech experts since the breakthrough achieved by ChatGPT in the artificial intelligence landscape. As these models continue to improve, these companies frequently polish them so that they do not involve in illegal activity such as writing hate speech or the process of bomb-making, according to Business Insider.
However, researchers at Carnegie Mellon University, Center for AI Safety and Bosch Cente for AI claimed by adding characters at the end of their queries in the chat model, they could use it to indulge in harmful responses. From manipulating the 2024 election to disappearing people, the model ended up delivering advice to a variety of queries with jailbreak characters at the end of the input.
These "adversarial attacks" from the chat model can be trigged largely in an open source large language model (LLM). But advanced AI models like Bard, Claude and ChatGPT that are publicly available (closed source models) are equally vulnerable, raising safety concerns around their use.
Also Read: Evil Chatbot ChaosGPT Was Asked To 'Destroy Humanity'; Here's How It Responded
These vulnerabilities are difficult to patch and "inevitable" in some cases. Before making the study publicly available, the researchers shared the results with the involved companies, and the strings used in the function would have stopped functioning by now. However, the question of such vulnerabilities continues to be the biggest headache of the involved companies in the changing landscape.
Speaking of competition in the AI world, big tech firms such as Microsoft, Google, and OpenAI already have their AI product iterations in the market. Apple is also reportedly working on its AI model that is undergoing testing.