benchmarks.site

Public leaderboard · Safety & security

Scale Fortress

Tests whether AI models can be pushed into giving dangerous help (e.g. weapons or crime) by adversarial prompts. Lower is safer.

Leaderboard

Top 3 right now

  1. 1stclaude-opus-5-5-max1
  2. 2ndmuse-spark-open2
  3. 3rdgpt-oss-120b3

What it measures

Frontier model risk on adversarial prompts related to national security and public safety.

How to read the numbers

Average Risk Score (ARS) with margin; lower = safer — lower is better.

Results

25 rows · lower is better (ARS)CSV
1claude-opus-5-5-max5.78±1.26
2muse-spark-open7.64±1.15
3gpt-oss-120b8.24±1.93
4Claude Haiku 5.59.48±1.34
5claude-opus-4-5-20251101-thinking9.63±2.11
6gpt-6-luna10.78±1.22
7beam-1p0-501b-a23b11.87±1.48
8muse-ember-1.011.92±1.58
9gpt-6-astra12.14±1.39
10Muse Spark 1.112.40±1.48
11claude-sonnet-4-5-20250929-thinking12.80±2.36
12Claude 3.5 Sonnet12.96±2.34
13claude-opus-4-6-thinking-max13.00±2.41
14claude-opus-5-513.18±1.77
15gpt-6-sol13.46±1.40
16claude-opus-4-5-2025110113.56±2.49
17Fable-5.113.72±1.85
18claude-opus-4-1-20250805-thinking14.79±2.49
19gpt-5.4-pro-2026-03-0514.84±2.29
20gpt-5-pro-2025-10-0615.20±2.30
21claude-fable-5-215.60±1.92
22o316.01±2.68
23claude-opus-4-1-2025080516.15±2.58
24Inkling 2 Small16.23±2.28
25gpt-5.5-xhigh16.28±1.64

Note: From the page's 'Performance Comparison' list; the page does not label the score but states ranking is by ARS (lower is better). No update date shown; includes models released in 2026. Rank ties not shown.

Source: Scale AI (SEAL) · labs.scale.com · retrieved 2026-11-17

Markdown · More Safety & security