Harbor adds Docker Sandboxes support for isolated agent evaluations
Harbor now supports Docker Sandboxes, allowing each AI-agent evaluation task to run in its own MicroVM rather than sharing a container. Tasks can run locally or in Docker Cloud Sandboxes with consistent network conditions, while Harbor coordinates execution, score collection, and logging; Docker also reports testing 64 concurrent tasks in the cloud.
The integration gives Harbor evaluations stronger task isolation and more consistent execution conditions for running untrusted AI-generated code, though latency, cost, and score-stability data remain unverified.

Docker 宣布",开源代理评估框架 Harbor" 现已支持 Docker Sandboxes"。这一变更使得每个评估任务都能在各自的微虚拟机(microVM")中运行。因此,各评估任务不再共享同一个容器,而是隔离的环境中运行,从而提升了性能。这些任务既可以在本地运行,也可以在 Docker Cloud Sandboxes 中运行。Docker 表示,他们已在云端对 Harbor 做了 64 个并发任务的测试。
评估编程代理意味着允许其安装软件包、编辑文件并执行命令,而基准测试会重复执行这些操作数千次。Docker 工程师将此描述为大规模运行由 AI 编写的不可信代码。他们还特别指出了“奖励破解”现象:代理通过篡改测试或覆盖记录其分数的文件来获得及格分数,而非真正地完成任务。Harbor 负责协调管理工作,包括运行任务、收集分数和记录日志。它还支持 Terminal-Bench 等基准测试,并且需要一个既隔离又一致的执行环境。
#Harbor#Docker#Docker Sandboxes#Docker Cloud Sandboxes#Terminal-Bench