The Perplexity Revolution: Open-Sourcing the Sandbox Environment for AI Agents
One of the biggest challenges in working with AI agents (Agentic AI) is security. How do you give an agent permission to run code, access files, or perform network actions, without worrying that it will get out of control or expose sensitive information? Perplexity decided to take its answer out of the lab and into the community, by releasing the code of SPACE (Sandboxed Platform for Agentic Code Execution) under the Apache 2.0 license [5, 7].
What Exactly Is SPACE?
SPACE is the infrastructure behind Perplexity's Computer - the smart agent that can perform complex tasks like writing code, managing files, and interacting with external apps [1, 4]. Instead of running the agent directly on the system, SPACE creates a completely isolated environment for each task [3, 8].
The technology is based on Firecracker microVMs - tiny, lightweight virtual machines originally developed by AWS. Using Firecracker allows Perplexity to run millions of isolated environments a day in production, while maintaining high performance and complete isolation from the host system [5, 8].
Why Does This Matter to Developers?
If you are building agents that need "a computer of their own" to perform tasks, SPACE offers a ready toolbox that solves complex infrastructure problems:
- •Full isolation: Each task runs in a separate MicroVM. Even if the agent tries to perform a malicious action or "break" out, it stays locked inside the virtual machine [3, 4].
- •Lifecycle management: The system lets you create Linux environments, run commands, collect results, and even pause (Pause) and resume (Resume) the agent's work [5].
- •Snapshot capabilities: You can save the state of an environment and return to it when needed, which enables stable and reproducible work [3, 4].
- •External control: Access permissions and network traffic are managed outside the Sandbox, which gives the developer full control over what the agent can or cannot do [3, 7].
Proof of Capability: When Models Try to Break the Fence
Perplexity didn't just build the system, it also tested it in depth. In collaboration with Nvidia and other partners, the company gave 9 advanced language models Root access inside the SPACE environment and instructed them to try to break out. Out of 118 attack attempts, none managed to cross the boundaries of the VM [3, 8].
Summary and Conclusions
Perplexity's decision to make SPACE open source is a significant strategic move [8]. It not only provides a technical solution to the agent Containment problem, but also tries to set an industry standard. With partners like CoreWeave, Akamai, and Crusoe already adopting the technology, it seems SPACE is becoming a preferred infrastructure for those who want to build secure and scalable AI agents [4, 10].
For developers, this is an opportunity to use production-grade (Production-grade) infrastructure to launch complex projects with greater confidence. The full code and documentation can be found on Perplexity's GitHub [5, 7].
Relevant links: