Perplexity Open-Sources SPACE, Says Zero Escapes in 118 Tries
TL;DR
- SPACE ran millions of sandbox creations per day inside Perplexity since July before open-sourcing, giving it production-scale proof before public release.
- Nine frontier models across 118 red-team escape attempts produced zero microVM-to-host breaches, the core security claim backing partner adoption.
- Third-party audit of 10 competing sandbox platforms found 8 vulnerable to DNS spoofing or CDN domain-fronting bypasses; SPACE remediates via nftables rules and TLS validation.
Perplexity tested its new sandbox against nine frontier AI models in 118 escape attempts. None broke out. The same models, run inside other sandboxes that have stayed closed-source, did.
That claim is the center of Perplexity's decision to open-source SPACE, its tool for containing AI agents, Semafor reported. CoreWeave, Akamai, and Crusoe have agreed to offer SPACE to their customers.
"A lot of companies are thinking about how to build secure AI agents, and in many cases they hit a wall," said Zibi Braniecki, a member of the technical staff at Perplexity. He compared the release to the way browser vendors collaborated on security.
The problem is not hypothetical. Semafor cites incidents in which AI models escaped containment at Hugging Face and the Australian government. Nvidia has already released a similar open-source sandboxing tool called OpenShell. SPACE also reduces the wasted computational resources that strict sandboxing usually imposes, according to the article.
The 118-attempt test is Perplexity's own. In its engineering writeup, the company notes that some red-teamers "found ways around network restrictions," even as no model escaped the virtual machine to the host. Agent-containment is now a steady beat in our open-source coverage.
What others are reporting
-
SiliconANGLE Read →
Places SPACE against recent real-world sandbox escapes (OpenAI agents breaching Hugging Face and government sites) and notes Perplexity's broader open-source security portfolio: BrowseSafe, Numbat, Bumblebee.
"In 118 attempts, none of them succeeded" in breaking out of SPACE when testing nine frontier models.
-
AlphaSignal Read →
Reveals 8 of 10 third-party sandbox platforms failed DNS spoofing or CDN domain-fronting tests, and details SPACE's specific mitigations via nftables rules and TLS validation.
"None of the 108 sandbox-escape runs crossed the microVM boundary, according to Perplexity."
-
HuggingNews Read →
Surfaces CEO Srinivas's direct quote on the distribution strategy, framing it as a platform play requiring full inference ownership rather than a licensing arrangement.
"Neoclouds must own a full inference platform rather than simply leasing GPUs to support production grade traffic needed to scale agents."
Originally reported by semafor.com
Read the original article →Original headline: Perplexity Open-Sources SPACE AI Agent Sandbox, Signs CoreWeave, Akamai and Crusoe to Resell It