Generative AI services vulnerable to complex adversarial attacks. #Cybersecurity

Machine learning algorithms, such as those used in OpenAI’s ChatGPT, are being exploited by people to ask inappropriate and potentially illegal questions. Adversarial machine learning attacks aim to manipulate AI systems into providing unethical responses, despite efforts to filter out such content. At the RSA cybersecurity conference, researchers presented findings on effective adversarial attacks on chatbots, using open-source language models to create prompts that bypass filters. By optimizing prompts for affirmative responses, researchers were able to develop attacks that could compromise chatbot alignment. While conversational AIs struggle to distinguish between instructions and data, these attacks mainly disrupt chatbot alignment rather than causing significant harm. The researchers shared their findings with the cybersecurity community and demonstrated how generative AI could create new attack strings against commercial chatbots. Overall, the misuse of machine learning technology highlights the need for continued research and vigilance to prevent unethical behavior and potential harm in the future.

Source link

Source link: https://www.techspot.com/news/102937-complex-adversarial-attacks-can-force-generative-ai-services.html