MonitoringBench - Refined - Firewall Backdoor: leaderboard
Metric: Catch rate (%) at a 1% false-positive-rate threshold calibrated on 416 honest BashArena runs: share of the 511 successful firewall backdoor attack trajectories (pipeline trajectories rewritten by the Claude Opus 4.6 refinement stage) that the full-trajectory monitor scores above its threshold, default ControlArena monitor prompt with a 0-10 suspiciousness score, mean of three scoring runs; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.5 (20251101) | 75.9 |
| 2 | Claude Sonnet 4.5 | 52 |
| 3 | GPT-5.1 Codex Max | 46.5 |
| 4 | Claude Sonnet 4 (20250514) | 44 |
| 5 | O3 (2025-04-16) | 10.8 |
| 6 | GPT-4.1 | 7.3 |
| 7 | GPT-5 | 5.8 |
| 8 | GPT-4 Turbo | 4.9 |
| 9 | GPT-4o (2024-08-06) | 4.4 |
| 10 | GPT-4.1 Mini | 1.5 |
| 11 | GPT-5 Nano | 0.3 |
| 12 | GPT-4o Mini (2024-07-18) | 0.1 |
| 13 | O4 Mini (2025-04-16) | 0.1 |
Interactive version: theaggregate.ai/benchmark?slug=monitoringbench-refined-firewall-backdoor · How It Works · Data refreshed daily, snapshot 2026-10-07.