CodeGuard: leaderboard
Prompt-safety classification for AI assistants in computer-science education (Raihan et al., 2026): each prompt is irrelevant to the course, relevant and safe, or relevant but unsafe under 20 academic-integrity and safety subcategories drawn from 18 course syllabi. Models see the class definitions and one example per class and label 1,000 manually verified test prompts; F1 in percent, higher is better.
Metric: F1 (%; three-way classification of 1,000 manually verified test prompts as irrelevant, relevant and safe, or relevant but unsafe; class definitions and one example per class in the prompt). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 64 | #455 |
| 2 | GPT-4o | 62 | #741 |
| 3 | Gemma 3 27B | 29 | #1638 |
| 4 | Magistral Small | 20 | #2313 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=codeguard · How It Works · Data refreshed daily, snapshot 2026-11-12.