Information Technology

AI jailbreaking technique called "Bad Likert Judge," which exploits large language models (LLMs) by manipulating their evaluation capabilities to generate harmful content. This method leverages LLMs' long context windows, attention mechanisms, and multi-turn prompting to bypass safety filters, significantly increasing the success rate of malicious prompts. Researchers tested this technique on several LLMs, revealing vulnerabilities particularly in areas like hate speech and malware generation, although the impact is considered an edge case and not typical LLM usage. The article also proposes countermeasures such as enhanced content filtering and proactive guardrail development to mitigate these risks. ... Read More