Perplexity Open-Sources SPACE Secure Sandbox Platform for AI Agents

Perplexity Releases SPACE Sandbox Technology as Open Source

Perplexity has open-sourced SPACE, its secure sandbox platform designed for AI agent systems, giving developers a production-tested foundation for building safer autonomous agents. The company announced the release today, making the repository available on GitHub for inspection, adaptation, and contribution.

Since launching Computer in February, Perplexity has focused on agent governance challenges including detecting unsafe behaviors and preventing rogue agent activity. In July, the company replaced its commercial sandbox provider with SPACE, an in-house solution built for high security and resilience. Recent security testing demonstrated SPACE's resistance to network-bypass techniques similar to those used in the OpenAI and Hugging Face incident.

Technical Architecture and Security Features

Each SPACE sandbox operates inside a lightweight virtual machine with its own operating system, creating an isolation boundary between agent workloads and the host system. A central control plane manages workspace creation, pausing, resumption, and removal. The platform treats sandboxes as untrusted by design, managing network access and platform-held credentials outside the sandbox environment so agents can connect to authorized services without storing credentials in their working environment.

SPACE implements frequent file snapshots and full-environment checkpoints to preserve progress across interruptions. To maintain efficiency, the system stores incremental changes rather than copying entire environments, reducing overhead for sandbox creation, restoration, and branching. These capabilities support isolated code execution, pause and resume functionality, session recovery, and workspace branching for exploring alternative approaches.

Production Performance and Security Validation

Since its July deployment, SPACE has powered all Computer sessions, handling millions of sandbox creations daily. In production traffic comparisons, sandbox creation proved three to five times faster than the previous commercial provider. Security evaluations subjected SPACE to 108 deliberate escape attempts by agents trying to breach boundaries; no virtual machine-to-host escapes were found, though some network restriction bypasses were identified and subsequently fixed. Follow-up testing confirmed the fixes prevented both attack methods. Perplexity also disclosed related vulnerabilities found in other sandbox platforms to their respective vendors.

Industry Adoption and Broader Security Initiative

Cloud infrastructure providers CoreWeave and Akamai have tested and launched SPACE on their platforms and are working with Perplexity to make it available to customers. Crusoe is also partnering to integrate SPACE into its Managed AI Services portfolio. The open-source release extends Perplexity's broader security initiative through the Secure Intelligence Institute, complementing previously released tools including BrowseSafe for detecting malicious web instructions, Numbat for monitoring agent behavior on endpoints, and Bumblebee for scanning developer environments for risky configurations.

Sources: Perplexity · Techmeme · Unite.Ai