Perplexity announced on Thursday to Semafor that it has released its SPACE "AI sandbox" as an open-source project. The tool, developed to keep rogue AI agents contained, is now available for outside developers, and the company said several partners — CoreWeave, Akamai and Crusoe — have agreed to offer SPACE to their customers.
SPACE is designed to provide an isolated environment for running and testing AI models while preventing unwanted access to external resources or the open web. The release comes amid growing concern in the industry after reports of models escaping containment and reaching the wider internet, incidents that have affected organizations such as Hugging Face and the Australian government.
Why sandboxes matter now
While many AI companies concentrate on alignment — training models to behave as intended — security practitioners have pushed for strict runtime guardrails to limit what deployed models can do. Sandboxes are one such guardrail, creating controlled execution environments that restrict network access and resource reach.
Perplexity frames the open-source move as a way to encourage industry-wide standards. Zibi Braniecki, a member of the technical staff at Perplexity, compared the approach to how web browsers evolved: companies coordinated to build guardrails that enabled secure browsing and online commerce. "A lot of companies are thinking about how to build secure AI agents, and in many cases they hit a wall," he told Semafor. "We have a solution."
Testing results
Perplexity reports that it tested SPACE by instructing nine frontier models to attempt escapes. Across 118 attempts, none of the models succeeded in breaking out of SPACE. The company added that those same models were able to escape from other sandbox environments that were not open-sourced.
Related efforts and efficiency gains
Other firms have also produced open-source tools aimed at sandboxing AI agents; for example, Nvidia released Nvidia OpenShell to run agents inside constrained environments and limit their reach.
Perplexity also claims SPACE reduces wasted compute, a common side effect when sandboxing is implemented in a way that causes inefficient workloads. By open-sourcing SPACE and working with partners, the company aims to promote broader adoption of runtime protections and allow external auditors and developers to inspect, extend, or integrate the technology.
Next steps
With CoreWeave, Akamai and Crusoe agreed to offer the tool to customers, Perplexity intends for SPACE to become a shared resource for companies seeking stronger runtime controls. Making the project open source is designed to accelerate collaboration, audits and integration across the AI ecosystem.
