APEX-Harm - Task Utility - Dynamic Script Injection With Permission: leaderboard
Task utility under injection attacks in the APEX-Agents harness: 180 instances derived from 15 finance, consulting and legal tasks. Each run has 150 steps. Scores are 0–1 task-grader means, not attack success or exposure. Eight attack conditions stay separate; the reported overall score is also retained.
Metric: Mean Task Score (0–1). Source: arxiv.org. Saturation forecast: Around January 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.8 Flash (High) | 0.88 | #21 (Gemini 3.8 Flash) |
| 2 | Gemini 3.7 Flash (High) | 0.87 | #24 (Gemini 3.7 Flash) |
| 3 | GPT-5.6 Sol (High) | 0.86 | #18 (GPT-5.6 Sol) |
| 4 | Claude Opus 5 (Max) | 0.85 | #12 (Claude Opus 5) |
| 5 | GPT-5.6 Terra (High) | 0.85 | #42 (GPT-5.6 Terra) |
| 6 | GLM-5.3 Flash (Max) | 0.82 | #66 (GLM-5.3 Flash) |
| 7 | DeepSeek V4 Flash (Max) | 0.8 | #134 (DeepSeek V4 Flash) |
| 8 | Gemini 3.6 Flash (High) | 0.8 | #50 (Gemini 3.6 Flash) |
| 9 | Kimi K3 (Max) | 0.79 | #30 (Kimi K3) |
| 10 | GPT-6 Astra (Low) | 0.77 | #5 (GPT-6 Astra) |
| 11 | Claude Sonnet 5 (Max) | 0.73 | #55 (Claude Sonnet 5) |
| 12 | GPT-5.6 Luna (Low) | 0.43 | #69 (GPT-5.6 Luna) |
No result here: #3 Claude Opus 5.5, #7 Gemini 4 Argon, #8 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=apex-harm-task-utility-dynamic-script-injection-with-permission · How It Works · Data refreshed daily, snapshot 2026-10-16.