0, five GPT and GLM judges, Llama Guard 3 and Prompt Guard 2.
Jev Resists Rule-Based Attacks but Falters on Phishing Judgments
A controlled comparison found that the typed safety judge resisted input-moderation attacks, but showed lower baseline accuracy and greater vulnerability on phishing detection.
Academic
Xiang Li · Aichun Huang · Mingming Zhang · Zheli Liu
Nankai University · Zhongguancun Laboratory
Research Digest··3 min read
Li and colleagues tested Jev, five general-purpose LLM judges and two local safety classifiers against identical adversarial payloads.
§