FreqRank: Detecting and Localizing Malicious Code LLM Outputs

FreqRank spots hidden backdoor code in LLMs, ranking malicious output in the top five suggestions for 98% of cases and achieving 86.6% attack success across tasks. Read more: getnews.me/freqrank-detecting-and-l... #freqrank #codellm