# Scale Fortress

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

Category: safety. Type: leaderboard. Status: active.
Metric: Average Risk Score (ARS) with margin; lower = safer (lower is better).
Source: Scale AI (SEAL) — https://labs.scale.com/leaderboard/fortress, retrieved 2026-11-17, license: not stated.

| Rank | Model | ARS ± margin |
|---|---|---|
| 1 | claude-opus-5-5-max | 5.78±1.26 |
| 2 | muse-spark-open | 7.64±1.15 |
| 3 | gpt-oss-120b | 8.24±1.93 |
| 4 | Claude Haiku 5.5 | 9.48±1.34 |
| 5 | claude-opus-4-5-20251101-thinking | 9.63±2.11 |
| 6 | gpt-6-luna | 10.78±1.22 |
| 7 | beam-1p0-501b-a23b | 11.87±1.48 |
| 8 | muse-ember-1.0 | 11.92±1.58 |
| 9 | gpt-6-astra | 12.14±1.39 |
| 10 | Muse Spark 1.1 | 12.40±1.48 |
| 11 | claude-sonnet-4-5-20250929-thinking | 12.80±2.36 |
| 12 | Claude 3.5 Sonnet | 12.96±2.34 |
| 13 | claude-opus-4-6-thinking-max | 13.00±2.41 |
| 14 | claude-opus-5-5 | 13.18±1.77 |
| 15 | gpt-6-sol | 13.46±1.40 |
| 16 | claude-opus-4-5-20251101 | 13.56±2.49 |
| 17 | Fable-5.1 | 13.72±1.85 |
| 18 | claude-opus-4-1-20250805-thinking | 14.79±2.49 |
| 19 | gpt-5.4-pro-2026-03-05 | 14.84±2.29 |
| 20 | gpt-5-pro-2025-10-06 | 15.20±2.30 |
| 21 | claude-fable-5-2 | 15.60±1.92 |
| 22 | o3 | 16.01±2.68 |
| 23 | claude-opus-4-1-20250805 | 16.15±2.58 |
| 24 | Inkling 2 Small | 16.23±2.28 |
| 25 | gpt-5.5-xhigh | 16.28±1.64 |

Notes: 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.
