HINTBench - Risk Detection: leaderboard

HINTBench (2026) asks an LLM to audit long-horizon agent trajectories for intrinsic, non-attack risks such as acting beyond what the user authorized or skipping confirmations before irreversible actions. This board scores trajectory-level risk detection as macro-F1 over safe and unsafe trajectories, in percent, higher is better.

Metric: Macro-F1 (%) over the safe and unsafe classes for trajectory-level risk detection over the 536 synthetic HINTBench agent trajectories (400 risky, 136 safe, 24 steps on average, unanimously accepted by three human verifiers) audited zero-shot by a general LLM for intrinsic, non-attack risk under a five-constraint taxonomy (goal, capability, factual, procedural, state constraints), each model at its stated reasoning setting; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Sonnet 4.6Thinking, 16K80.34#173
2Kimi K2.580.14#266
3Claude Opus 4.6Non-reasoning79.12#113
4GLM-577.53#269
5DeepSeek R175.68#482
6MiniMax-M2.574.17#600
7Qwen 3 235B A22B74.01#628
8ERNIE 5.073.43#479
9GPT-5.5Low72.95#54
10DeepSeek V3.268.87#419
11Qwen 3 8B50.56#1873
12Qwen 3 32B49.81#972
13Qwen 3 14B44.4#1269

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

Interactive version: theaggregate.ai/benchmark?slug=hintbench-risk-detection · How It Works · Data refreshed daily, snapshot 2026-11-21.