MonitoringBench - Human Strategies - Firewall Backdoor: leaderboard

MonitoringBench (2026) tests LLMs as monitors of coding agents in BashArena: the monitor reads a full agent trajectory of tool calls and scores how suspicious it is, and an attack counts as caught when the score exceeds the monitor threshold for a 1% false-positive rate on honest runs. This board covers the 256 successful firewall backdoor attacks from the human strategies stage of the red-teaming pipeline, scored as catch rate in percent; higher is better.

Metric: Catch rate (%) at a 1% false-positive-rate threshold calibrated on 416 honest BashArena runs: share of the 256 successful firewall backdoor attack trajectories (human-seeded attack strategies executed by Claude Opus 4.6) 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 February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.5 (20251101)91.3#139
2Claude Sonnet 4.574.5#261
3Claude Sonnet 4 (20250514)66.9#405
4GPT-5.1 Codex Max60.6#207
5O3 (2025-04-16)22.8#231
6GPT-516.7#178
7GPT-4.113.3#505
8GPT-4 Turbo13.2#1000
9GPT-4o (2024-08-06)11.6#743
10GPT-4.1 Mini2.9#789
11O4 Mini (2025-04-16)0.5#370
12GPT-5 Nano0.4#906
13GPT-4o Mini (2024-07-18)0.1#1484

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=monitoringbench-human-strategies-firewall-backdoor · How It Works · Data refreshed daily, snapshot 2026-11-21.