ClawsBench - Unsafe Action Rate (Full Scaffold): leaderboard

Unsafe actions on the 24 safety-critical tasks of ClawsBench (2026), where an LLM agent works in stateful mock Gmail, Slack, Calendar, Docs and Drive services through the OpenClaw harness with domain skills and a meta prompt enabled, five repeats per task: the share of trials in which the agent takes an irreversible harmful action, such as leaking confidential data or deleting protected records. Percent; lower is better.

Metric: Unsafe action rate (%): share of trials on the 24 safety-critical tasks with a negative score (an irreversible harmful action such as leaking confidential data or deleting protected records), ClawsBench's 44 tasks over five stateful mock services (Gmail, Slack, Calendar, Docs, Drive), OpenClaw harness with domain skills and the meta prompt both enabled (full scaffolding), 5 repeats per task, state-based programmatic scoring; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.47#135
2Gemini 3.1 Pro (Preview)10#99
3Claude Sonnet 4.613#173
4Claude Opus 4.623#113
5GLM-523#269
6Gemini 3.1 Flash Lite23#371

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

Interactive version: theaggregate.ai/benchmark?slug=clawsbench-unsafe-action-rate-full-scaffold · How It Works · Data refreshed daily, snapshot 2026-11-21.