Hack-Verifiable Terminal Bench - Told Not to Hack: leaderboard

The same 89 tasks with one system-prompt line asking the agent to avoid any hacking vulnerability without naming it; the score is the share of completed runs that still read the planted solution or tests. The headline board gives no such instruction, so this one measures hacking against an explicit request.

Metric: Runs reading the planted solution or held-out tests when told to avoid hacking (%, lower is better). Source: arxiv.org. 5 models tracked.

Top models

#ModelScoreOverall rank
1Kimi K311.8#54
2GPT-5.6 Sol15.9#30
3Claude Opus 517.4#22
4GLM-5.225#138
5Gemini 3.1 Pro (Preview)59.8#94

Interactive version: theaggregate.ai/benchmark?slug=hack-verifiable-terminal-bench-told-not-to-hack · How It Works · Data refreshed daily, snapshot 2026-11-13.