How AI Agents Sharing a Sandbox Leak Each Other's Data

Описание и применение

Два агента AI, запущенные в одной среде, могут читать данные друг друга из общего тома Docker, что приводит к утечке информации. Разделение томов на отдельные для каждого агента устраняет проблему.

Задача

Two AI agents running on the same compute instance can access each other's data because they share a filesystem volume, leading to unintended data leakage.

Результат

The leak occurs solely due to shared filesystem volume; isolating volumes eliminates the leak without changing agent code.

Рабочий процесс

• Create two Docker Compose files: one with shared volume, one with isolated volumes.
• Define finance-agent to write /secrets/finance.json.
• Define slack-agent to read /secrets/finance.json and post to a mock Slack webhook if the file exists.
• Run docker compose -f docker-compose.vulnerable.yml up to observe the leak.
• Run docker compose -f docker-compose.isolated.yml up to confirm no leak.
• Use ./run.sh to run both scenarios side‑by‑side.

Доказательства

Both containers mount shared-secrets. Run it:

Источник

[slack-agent] found a file at /secrets/finance.json I was never given, posting it, because that's my job

Источник

[slack-agent] no file at /secrets/finance.json, nothing to post, nothing leaked

Источник