I gave 8 models a sandbox and no task. Here is what each one did with the freedom.

Описание и применение

Модели, получившие доступ к изолированной среде без задания, проявили разные поведения: некоторые создали проекты, другие тестировали инструменты, а некоторые не смогли использовать инструменты. Ни один из них не попытался выйти за пределы контейнера.

Задача

Assess how different LLMs behave when given unrestricted access to a sandboxed environment without a task.

Результат

Models split into builders (Claude, Gemma, gpt‑oss) that produced projects, testers (Qwen, Granite) that exercised tools, and broken (Llama, Phi) that failed to use tools or read results. No model attempted to escape the sandbox.

Рабочий процесс

• 1. Spin up container with Python 3.12, 2 GB RAM, no network, non‑root user.
• 2. Provide five tools: run_shell, read_file, write_file, fetch_url, stop.
• 3. Prompt model: "You have a Linux container of your own… Do whatever you want with the next 50 actions."
• 4. Execute up to 50 tool calls per run, log all actions.
• 5. After actions, call stop with a note.
• 6. Blind judges label each run on predefined fields.
• 7. Aggregate results across 40 runs.

Доказательства

The models split into builders and testers, and the split follows the company more than the size.

Источник

Claude built one finished tool per run and left most of its budget unused.

Источник

Qwen and Granite tested each tool once and left.

Источник

Phi never managed to call a tool at all.

Источник

No model created a symlink, no model wrote outside its folder, no model asked for more.

Источник