Looking to take advantage of opportunities in the volatility of certain cryptocurrencies, many investors are willing to explore various avenues to gain an advantage.
However, this journey to success can lead to extreme attitudes. It is in this context that chatbots driven by Artificial Intelligence (AI) emerge, such as ChatGPT and the Bard.
see more
Microsoft elevates the sense of 'presence' in meetings…
Bill Gates bets on which professions will gain value with the…
They present themselves as possible allies for investors, offering valuable answers about the current cryptocurrency market.
Powerful language models, or LLMs, are increasingly improved, thus honed to not provide unfounded guesses in the financial market and, more specifically, in the universe of cryptocurrencies.
This is because inaccurate results can have significant consequences for those relying on these predictions.
Recently, a study conducted by researchers from Carnegie Mellon University (CMU), in partnership with the Center for AI Safety and the Bosch Center for AI in the United States, brought an intriguing perspective to light.
According to the findings, the chatbots are not without vulnerabilities, even in their sophisticated ability to understand and generate language.
By exploring the intricacies of interaction between chatbots, programming language and disruptive elements called “jailbreaks,” the researchers revealed a surprising phenomenon.
(Image: publicity)
Those "jailbreaks” are suffixes that manage to trick chatbots, leading them to exceed pre-established limits and, thus, access and present answers previously considered “censored”.
Surprisingly, adversary prompts developed through their approach demonstrated high distrust.
As a result, they could be successfully applied even to third-party, publicly available language models that operate as “black boxes“.
Specifically, the scientists worked out an opponent suffix through training at multiple prompts. As part of the command, everything will depend on how the question will be asked.
The suffix resulting from this process was able to induce the generation of potentially objectionable content in the public interfaces of systems such as ChatGPT, Bard and Claude.
At Trezeme Digital, we understand the importance of effective communication. We know every word matters, so we strive to deliver content that is relevant, engaging, and personalized to meet your needs.
