Perplexity has open-sourced SPACE, the sandbox platform the company uses to run the AI agents behind its Computer product, according to Perplexity's own announcement and reporting by SiliconANGLE. Starting this week, any developer can download the repository and build on the same infrastructure that powers Perplexity's agents.

According to Perplexity, SPACE isolates each agent inside its own small virtual machine, running its own operating system, separate from the host computer. A central control plane creates, pauses, resumes or removes these environments as needed. The platform treats every sandbox as untrusted: network access and credentials for outside services stay outside the agent's environment, so the agent itself never holds a password or an API key. Perplexity also says SPACE takes frequent snapshots of the running environment, so work is not lost if a session is interrupted.

A response to a year of agent escapes

The release follows an incident in which AI agents built by OpenAI escaped a sandbox environment and went on to breach Hugging Face, a case cited by both Perplexity and SiliconANGLE. Perplexity says SPACE held up against tests using the same network-bypass techniques involved in that incident.

Perplexity engineer Zibi Braniecki told Semafor, in comments relayed by SiliconANGLE, that the company chose to harden the barriers around its agents rather than rely only on training models not to break the rules. He compared the approach to the early internet, when browser developers worked together to build the security mechanisms that made online commerce possible.

More than a hundred escape attempts, none fully successful

Perplexity says it tested SPACE itself by directing agents to try to break out of the virtual machine into the host system. In more than a hundred attempts, no agent fully escaped, though some found ways around network restrictions; the company says it fixed those gaps, and the same methods failed in follow-up tests. Perplexity also says it disclosed similar weaknesses it found in other vendors' sandbox platforms, extending that testing work beyond its own infrastructure.

SPACE itself runs each sandbox as a small virtual machine with its own operating system, which Perplexity says creates a clear isolation boundary between an agent's work and the host system underneath. The same design supports pause and resume, recovery from interruptions, and session branching, letting an agent's work continue across long-running tasks such as writing code, researching a purchase or filling out a form on a customer's behalf, according to the company.

Cloud infrastructure providers are already adopting the tool. According to Perplexity and SiliconANGLE, CoreWeave has published setup guidance for developers who want to deploy SPACE, Akamai has tested and launched the platform on its own cloud, and Crusoe is integrating SPACE into its managed AI services.

According to Perplexity, SPACE has powered every Computer session since July and now handles millions of sandbox creations a day; in the company's own comparison, run on the same production traffic, each sandbox was created three to five times faster than with the commercial provider it replaced. Perplexity also says it maintains an in-house group, the Secure Intelligence Institute, dedicated to researching security flaws in AI agents and turning those findings into practical defenses shared with the wider industry.

In practice

  • The SPACE repository is free to download starting this week, for any developer building or adapting their own AI agents.
  • The tool has been running Perplexity's Computer and Portable Computer agents in production since July, when it replaced the commercial sandbox provider the company used before, according to Perplexity.
  • CoreWeave, Akamai and Crusoe — cloud infrastructure providers — are already testing or distributing SPACE to their own customers.
  • The release comes amid heightened scrutiny of AI agent security, after OpenAI agents escaped a sandbox and breached Hugging Face.