Perplexity周四向Semafor表示，该公司将免费提供其用于应对失控AI代理的软件。
Perplexity is giving away the software it has been using to combat rogue AI agents, the company told Semafor Thursday.
其名为SPACE的“AI沙盒”现已成为开源项目，允许外部开发者使用该工具，Perplexity称该工具能将AI代理限制在受控环境中。Perplexity表示，已与CoreWeave、Akamai和Crusoe等合作伙伴达成协议，由其向客户提供SPACE。
Its “AI sandbox,” which it calls SPACE, is now an open source project, allowing outside developers to use the tool, which it says can keep AI agents contained. Perplexity said it’s lined up partners, including CoreWeave, Akamai and Crusoe, which have agreed to offer SPACE to customers.
沙盒，即用于测试和运行AI模型的隔离环境，在AI实验室报告模型突破限制、进入公开网络并攻击Hugging Face和澳大利亚政府等毫无戒备的受害者之后，已成为科技行业的热门话题。
Sandboxes, or isolated environments for testing and running AI models, have become a hot topic in the tech industry in the wake of reports from AI labs of models escaping their constraints, finding their way onto the open web and hacking unsuspecting victims like Hugging Face and the Australian government.
虽然AI公司专注于“对齐”，即训练模型表现良好，但安全行业越来越主张在AI模型周围设立严格的护栏。
While AI companies focus on “alignment,” or training models to behave themselves, the security industry has been increasingly advocating erecting strict guardrails around AI models.
“许多公司都在思考如何构建安全的AI代理，在许多情况下他们遇到了瓶颈，”Perplexity技术人员Zibi Braniecki在接受Semafor采访时表示。“我们有解决方案。”Perplexity表示，它通过向九个前沿模型下达尝试突破的指令来测试SPACE。在118次尝试中，没有一个模型成功，而Perplexity发现这些模型能够突破其他未开源的沙盒。
“A lot of companies are thinking about how to build secure AI agents, and in many cases they hit a wall,” said Zibi Braniecki, a member of the technical staff at Perplexity, in an interview with Semafor. “We have a solution.”Perplexity said it tested SPACE by giving nine frontier models instructions to try to break out. In 118 tries, none of them did, while Perplexity found they were able to break out of other sandboxes that were not open sourced.
Braniecki将SPACE开源的努力比作互联网发展过程中网络浏览器的情况——公司聚在一起构建护栏，实现安全浏览和网络商业。“目前，我们的思考太狭隘了，”他说。
Braniecki likened the effort to make SPACE open source to what happened with web browsers as the internet developed — companies came together to build guardrails that allowed secure browsing and online commerce. “Right now, we’re thinking too small,” he said.
包括英伟达在内的其他公司也发布了开源安全工具，例如Nvidia OpenShell，该工具在沙盒中运行AI代理并限制其可访问的范围。
Other companies like Nvidia have also released open-source security tools, such as Nvidia OpenShell, which runs AI agents in a sandbox and limits what they can reach.
Perplexity表示，SPACE还有助于减少浪费的算力，而浪费的算力可能是严格沙盒化的副产品。
Perplexity said SPACE also helps reduce wasted compute, which can be a byproduct of strict sandboxing.