Perplexity has officially open-sourced SPACE, its in-house agent sandbox platform designed to handle secure, high-volume code execution for AI assistants. Released under the Apache License 2.0, the platform is now available with a public repository and comprehensive documentation. According to the company, SPACE was developed to replace an external commercial sandbox provider and has been powering all Perplexity Computer sessions since July, managing millions of sandbox creations daily.
Architecture and Execution Mechanics
At its core, SPACE—which stands for Sandboxed Platform for Agentic Code Execution—operates by running code inside isolated Firecracker microVMs. Rather than embedding a heavy Rust library, applications interact with the platform through an asynchronous HTTP API and server-sent events. This design allows developers to programmatically create Linux environments, execute commands, collect outputs, and manage complete sandbox lifecycles without direct language lock-in.
The platform relies on templates to provision starting environments, giving each active sandbox its own dedicated virtual machine and writable filesystem. To optimize performance during heavy operational workloads, SPACE avoids copying entire environments repeatedly. Instead, it maintains frequent file snapshots and checkpoints, storing only incremental changes to preserve running state and reduce the overhead associated with branching or restoring workspaces. Additionally, the project includes compatibility layers such as an E2B wrapper designed to work with standard Python SDKs.
Security Controls and Cloud Integration
Security and isolation form the backbone of the SPACE architecture. Each sandbox functions within a strict virtual machine boundary, isolating agent workflows from the underlying host system. Furthermore, sensitive platform credentials and network access routes are managed entirely outside the working environment by a central control plane. This separation ensures that autonomous agents can interact with authorized external services without storing sensitive authentication tokens locally.
Perplexity reports that SPACE has undergone rigorous internal security evaluations, demonstrating resilience against network-bypass techniques during adversarial testing. The platform joins a broader portfolio of defensive tools released through Perplexity’s Secure Intelligence Institute, sitting alongside utilities like BrowseSafe, Numbat, and Bumblebee. Beyond internal testing, cloud providers including CoreWeave, Akamai, and Crusoe have evaluated the platform and are actively working to incorporate SPACE into their respective managed services.
Strategic Implications for AI Agent Development
The decision to open-source SPACE highlights a critical shift in how the AI industry approaches agent reliability and runtime governance. As autonomous agents move from simple chatbots to active participants capable of executing code, manipulating files, and interacting with local environments, the attack surface expands exponentially. Traditional web application sandboxing is often inadequate for unpredictable agent behaviors, making robust container and microVM isolation a mandatory architectural layer for modern development teams.
By releasing SPACE under an open-source license, Perplexity addresses a major industry bottleneck: the lack of production-tested, standardized infrastructure for agent execution. Developers building autonomous systems frequently struggle with state recovery, session pausing for human-in-the-loop verification, and multi-workspace branching. Providing a battle-tested platform that handles these edge cases natively allows engineering teams to focus on core application logic rather than spending valuable resources building custom isolation layers from scratch. Moreover, open-sourcing the runtime invites broader community scrutiny, which is essential for identifying subtle security vulnerabilities and hardening agent environments against sophisticated jailbreak and escape tactics.
Cover: AI-generated illustration with headline text; not a documentary photograph. Logo: Perplexity, via Wikimedia Commons (Public domain).
Source: Perplexity Open-Sources SPACE, Its In-House Agent Sandbox Platform · GitHub – perplexityai/SPACE: Sandboxed Platform for Agentic Code Execution: pluggable-runtime sandboxes for running agent and user code in isolation. · GitHub
SIGNAL TITANS
Leave a Reply