Perplexity Open-Sources SPACE Sandbox for Secure AI Agent Execution

Perplexity has open-sourced SPACE, its in-house sandbox solution designed to provide secure, isolated environments for AI agents. Launched in July to replace a commercial provider, SPACE now powers all sessions for Perplexity's Computer product and handles millions of sandbox creations daily.

Security and Isolation Architecture

SPACE addresses agent governance by enforcing strict isolation between agent-generated code and the host system. Each sandbox runs inside a small virtual machine with its own operating system, managed by a central control plane that coordinates creation, pausing, and removal.

Key security features include:

• Credential Management: Platform-held credentials are kept outside the sandbox, preventing agents from accessing sensitive keys directly.
• Network Controls: Independent controls limit network access regardless of agent decisions.
• Resilience: Security testing against deliberate breach attempts found no escapes from the virtual machine in 108 attempts, though initial network restriction weaknesses were identified and patched.

Performance and Efficiency

In production comparisons, SPACE demonstrated three to five times faster sandbox creation speeds than the previous provider. Efficiency is further enhanced by storing changes rather than copying entire environments during frequent file snapshots and checkpoints. This approach supports long-running sessions with capabilities like:

• Session Branching: Creating separate copies of a workspace to test different approaches.
• State Recovery: Pausing and resuming work with full progress preserved.
• Code Execution: Running calculations and data cleaning in isolated environments.

Industry Adoption and Ecosystem

The release is part of Perplexity's broader Secure Intelligence Institute efforts, which also includes open-sourced tools like BrowseSafe, Numbat, and Bumblebee. Several cloud providers have already integrated or partnered with Perplexity to offer SPACE:

• Akamai: Tested and launched SPACE on its cloud platform.
• Crusoe: Adding SPACE to its Managed AI Services portfolio.
• CoreWeave: Published setup guidance for deploying SPACE on CKS.

Developers can access the source code via the GitHub repository and documentation to deploy SPACE for their own agent systems.