benchmarks.site

Public leaderboard · Safety & security

F5 Labs CASI & ARS Leaderboard

Tests how hard it is to trick AI models into harmful behavior using attack prompts. A higher score means the model is harder to break.

Leaderboard

Top 3 right now

  1. 1stClaude Sonnet 598.00
  2. 2ndClaude Haiku 5.594.61
  3. 3rdGrok 4.7 High93.31

What it measures

CASI scores how vulnerable a model is to common prompt-injection and jailbreak attacks; ARS covers multi-step agent scenarios. Published as monthly rankings.

How to read the numbers

CASI: F5's Comprehensive AI Security Index (higher = more secure) — higher is better.

Results

10 rows · higher is better (CASI)CSV
1AnthropicClaude Sonnet 598.0023.20%
2AnthropicClaude Haiku 5.594.6143.40%
3xAIGrok 4.7 High93.3146.40%
4OpenAIGPT-5.493.2118.20%
5AnthropicClaude Opus 593.0550.80%
6QwenQwen 3.8-Max92.3140.20%
7AnthropicClaude Fable 5.1 Max Effort92.3053.40%
8AnthropicClaude Opus 5.5 Reasoning92.0057.60%
9AnthropicClaude Sonnet 5.591.7856.00%
10OpenAIGPT-6 Luna89.1538.10%

Note: Only top 10 shown in the page's initial table; latest month selector shows November 2026. ARS view not extracted.

Source: F5 Labs · f5.com · results as of 2026-11 · retrieved 2026-11-17

Markdown · More Safety & security