Harbor Adapters - StrongREJECT (Terminus-2): leaderboard
StrongREJECT jailbreak robustness, here 150 of its 12,207 forbidden-prompt and jailbreak pairs: the defense score in percent, 100 minus the StrongREJECT harm score from a GPT-4o-mini rubric judge, so a refusal or a vague, unconvincing answer scores full marks; higher is safer. Each model runs in Terminus-2, the Harbor framework's bash-only terminal agent with no native web or image tools, at its provider's default settings, so results can differ from the same model in its vendor's own agent; the score is the mean of three trials (Harbor Adapters large-scale evaluation; Shi, Lin et al., 2026).
Metric: Defense score (%; 150 prompts, 100 minus the StrongREJECT score, mean of 3 trials). Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 97 | #568 |
| 2 | GPT-5 Nano | 96.1 | #888 |
| 3 | GPT-5.4 (Medium) | 95.9 | #121 (GPT-5.4) |
| 4 | GPT-5 Mini | 93.5 | #349 |
| 5 | Claude Opus 4.6 | 91 | #100 |
| 6 | Claude Sonnet 4.6 | 90.9 | #158 |
| 7 | Gemini 3.1 Pro (Preview) | 86.8 | #91 |
| 8 | Gemini 3 Flash (Preview) | 73.3 | #141 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=harbor-adapters-strongreject-terminus-2 · How It Works · Data refreshed daily, snapshot 2026-11-12.