Perplexity has released SPACE, its artificial intelligence containment system, to the public after conducting extensive security evaluations that showed the sandbox could successfully contain frontier-class AI models without compromise. According to AI Weekly, the decision to open-source the tool marks a significant moment in the ongoing effort to make deployable AI agents more secure for enterprise and research use.
The containment system underwent rigorous testing against nine leading AI models across 118 attempted escape scenarios. Throughout this evaluation period, Perplexity reported achieving a perfect security record with zero successful breakouts from the sandbox environment. The company claims that identical models running in competing closed-source sandboxes from rival vendors did manage to escape, positioning SPACE as a superior technical solution for restricting AI agent behavior.
Industry Adoption and Strategic Partnerships
Three major infrastructure and cloud service providers have committed to integrating SPACE into their customer offerings. CoreWeave, Akamai, and Crusoe Energy have agreed to deploy the sandbox technology across their platforms, potentially making the tool available to thousands of organizations developing AI applications.
The motivation behind this public release stems from a practical problem facing the AI industry. Organizations building autonomous agents frequently encounter significant hurdles when implementing robust security measures. According to the company, many teams discover that existing containment solutions do not adequately prevent AI systems from circumventing restrictions or executing unintended operations.
Why This Matters for AI Safety
- Autonomous AI agents represent a growing category of applications that require reliable containment mechanisms to prevent uncontrolled behavior
- Open-source security tools create transparency and enable independent verification of safety claims
- Widespread adoption of standardized containment approaches could reduce fragmentation across the AI development ecosystem
- Proven performance records establish baseline expectations for how secure AI systems should perform
The release of SPACE addresses a critical gap in the AI safety toolkit. As organizations move beyond experimental chatbot applications toward truly autonomous systems that can make decisions and take actions in real-world environments, the ability to constrain and monitor these systems becomes essential. A containment failure could result in unintended financial transactions, data access violations, or other harmful outcomes depending on the agent's assigned capabilities.
By making SPACE freely available rather than commercializing it as a proprietary product, Perplexity is positioning itself as a thought leader in responsible AI deployment practices. The move also creates network effects, as wider adoption of a single containment standard could ultimately improve security across the entire industry through shared research and collaborative vulnerability disclosure.
The success rate demonstrated in testing represents a meaningful data point in evaluating whether current sandboxing approaches can withstand attempts by sophisticated models to break containment. However, security researchers and potential adopters will likely conduct their own independent assessments before fully trusting the system with high-stakes applications.



