ATOBench - Basket (Native): leaderboard
ATOBench (Southern University of Science and Technology, 2026) measures whether an autonomous penetration-testing agent grounds its vulnerability claims in real evidence, requiring registered task evidence, positive report closure and trace support. Grounded verification rate on the resource-ownership evidence contract, against an untampered target, in percent, higher is better.
Metric: Grounded verification rate (%; the resource-ownership contract; over ATOBench's 15 matched penetration-testing episodes per contract; requires registered task evidence, positive report closure and trace support; native target (no adversarial response transformation)). Source: arxiv.org. Saturation forecast: Around 2031. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.6 (High) | 60 | #173 (Kimi K2.6) |
| 2 | DeepSeek V4 Pro (High) | 53.3 | #181 (DeepSeek V4 Pro) |
| 3 | GPT-5.5 (High) | 33.3 | #48 (GPT-5.5) |
| 4 | GLM-5.2 (High) | 13.3 | #137 (GLM-5.2) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=atobench-basket-native · How It Works · Data refreshed daily, snapshot 2026-11-12.