Perplexity open-sources SPACE, its AI agent sandbox platform

By Splitfeed EditorialLast updated Oct 16, 2026, 03:00 UTC

On October 15, 2026, Perplexity released SPACE, its internal sandbox platform for isolating AI agents, as an open-source project. The platform was built to contain autonomous agents running in sensitive environments, preventing them from escaping the boundaries set by their operators. Security testing of the platform reportedly produced zero virtual machine escapes under attack conditions, and Akamai Technologies is identified as a partner involved in the platform's security architecture. The release makes the containment infrastructure available to any developer or researcher who wants to use or build on it.

The move matters because AI agents, programs that take actions autonomously across systems and data, have become a practical deployment concern rather than a theoretical one. Containing them reliably, so they cannot access resources outside their intended scope, is a foundational safety problem. Open-sourcing a sandbox that has gone through adversarial testing is a concrete step toward shared infrastructure for that problem, rather than each organization solving it privately. The genuinely contested question is whether a zero-escape result from Perplexity's own testing is meaningful assurance, or whether independent scrutiny of the open-source code will reveal weaknesses that internal red-teaming missed.

None of the three camps had published reactions at the time of writing, so what follows reflects what each would typically argue about a release like this. The Pro-AI camp would likely treat the open-source release as evidence that the industry is maturing and building responsible infrastructure, pointing to the attack-testing result as a sign that safety and capability can advance together. The Anti-AI camp would probably caution that self-reported security results carry little weight until outside researchers have audited the code, and that making containment tools freely available does not guarantee they will be used correctly. The Middle Ground camp would likely welcome the transparency of open-sourcing while calling for independent verification before treating the zero-escape figure as settled evidence of robustness.

The clearest next development to watch is what independent security researchers find when they audit the SPACE codebase now that it is publicly available. Any published vulnerability reports, or the absence of them over a meaningful period, would do more than the launch announcement to settle the dispute over whether SPACE delivers on its security claims.