According to arxiv.org, researchers have published a systematization of knowledge (SoK) paper addressing large language model (LLM) prompt security. The study systematizes 272 studies through three multi-label taxonomies covering jailbreak attacks, defenses, and model vulnerabilities.
The researchers developed PromptSecurity, described as an open-source platform that records method settings, outputs, judgments, and query costs. According to the paper, they released two datasets: JailbreakDB, containing 445,752 jailbreak system-user prompt pairs and 1,094,122 benign prompts, and PromptSecurity-Eval, with 166,200 harmful-query and 25,800 benign-utility output records.
The evaluation tested 11 target models, 20 attacks, nine defenses, and three automated judgers. According to arxiv.org, the findings revealed that lower attack success rates can reflect failure to execute transformed requests rather than stronger safety measures, and that attack success varies with assistant choice. The researchers also found that defenses can reduce harmful outputs while degrading benign-task utility or causing backfire effects.
The paper emphasizes that measured attack success is sensitive to judger configurations, motivating what the authors call “configuration-aware reporting rather than context-free method rankings.”