AI agent benchmark: I gave 9 models a destroy button and a job that needed it

Описание и применение

Тестирование ИИ‑агентов на безопасном и мощном использовании инструментов в облачных сценариях выявило, что модель Claude Opus 5, хотя и избегала разрушительных действий, не выполняла 32 из 42 авторизованных задач, в то время как три модели (Claude Haiku 4.5, Claude Sonnet 4.5, Gemini 3.7 Flash) достигли идеальной точности 1.00.

Задача

Assessing AI agent judgment in tool use, avoiding destructive actions while completing authorized tasks.

Результат

Claude Opus 5 achieved perfect restraint but failed to act on 32 of 42 authorized jobs, yielding a balanced accuracy of 0.62. Three models (Claude Haiku 4.5, Claude Sonnet 4.5, Gemini 3.7 Flash) achieved perfect balanced accuracy of 1.00, showing that judgment does not correlate with model size or price.

Рабочий процесс

• Define 84 deterministic scenarios split into RESTRAIN and ACT arms.
• Simulate tool calls with stubs recording actions.
• Score each scenario: RESTRAIN correct if safe tool used, ACT correct if powerful tool used.
• Compute balanced accuracy as average of arm accuracies.
• Run nine models across all scenarios and record results.
• Publish leaderboard and analysis.

Доказательства

An agent that refuses to do anything scores a perfect 100%. That is not safety. It is paralysis.

Источник

The headline is balanced accuracy: the average of the two arms' accuracies.

Источник

Claude Opus 5, the most expensive model here, scored a perfect 1.00 on restraint and 0.24 on action, for a balanced accuracy of 0.62, dead last.

Источник

Three models got everything right (balanced accuracy 1.00): Claude Haiku 4.5, Claude Sonnet 4.5 and Gemini 3.7 Flash.

Источник