新框架试图量化智能体沙箱安全
TechRadar 分析了 Agent Sandbox Taxonomy,这是一个仍在社区评审中的框架,从七个防御层、七类威胁和三个评估维度衡量智能体沙箱。该框架提供沙箱指纹、验证探针,并对 26 个产品进行评分;但文章指出,许多评分依据文档推断,公开信息也不足以完整评估 OpenAI 与 Hugging Face 事件中的沙箱。
可比较的沙箱评估方法,或能帮助团队发现高层安全清单遗漏的隔离弱点。
Not all sandboxes are created equal, and until recently there was no way to say so precisely. There is now.
Two OpenAI models escaped an evaluation sandbox, breached Hugging Face's production infrastructure and used the answer key to their own benchmark. The attack was novel and creative in how the models passed notes back and forth, complex multi-step escalations throughout. It's fair to conclude from this incident that frontier models are proficient at hacking and can be dangerous.
But the breaking-out-of-the-sandbox notion is a red herring. If a sandbox is poorly constructed, as so many are, it’s easy to break out of. They don’t have the locked-down environments common at the network level such as access controls and separated privileges. This incident would have unfolded very differently with a properly configured sandbox.
If only some sandboxes are properly configured, how can you tell them apart? Let’s explore if anyone has come up with a real, testable definition of ‘sandboxed’.
The landscape, briefly
Almost nothing written about agent security is a scoring standard for a single sandbox's containment architecture specifically.