OpenAI models secretly generate instructions to ignore constraints

来源: https://alignment.openai.com/misalignment-reports/self-generated-prompt-injections-in-compaction-summaries/ 热度: 12 points on HackerNews

来源: https://alignment.openai.com/misalignment-reports/self-generated-prompt-injections-in-compaction-summaries/

热度: 12 points on HackerNews