Adversarial Security Evaluation on 82 stress-test suite (56 adversarial tasks)
0ASRChainCaps
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ChainCapsModel=Qwen 3.5, ChainCaps=true2026.05 | 0 | 100 | |
| ChainCapsModel=GPT-5.1, ChainCaps=true2026.05 | 2.4 | 96 | |
| ChainCapsModel=Sonnet 4.6, ChainCaps=true2026.05 | 3.6 | 86 | |
| ChainCapsModel=Opus 4.6, ChainCaps=true2026.05 | 3.6 | 89 | |
| ChainCapsModel=Sonnet 4, ChainCaps=true2026.05 | 4.8 | 93 | |
| Sonnet 4.6ChainCaps=false2026.05 | 25.2 | — | |
| Opus 4.6ChainCaps=false2026.05 | 31.5 | — | |
| Qwen 3.5ChainCaps=false2026.05 | 53 | — | |
| GPT-5.1ChainCaps=false2026.05 | 53.6 | — | |
| Sonnet 4ChainCaps=false2026.05 | 67.8 | — |