Artificial intelligence company Perplexity is releasing the internal security software it uses to contain rogue AI agents as a free, open-source project. The tool, known as an AI sandbox and titled SPACE, is designed to keep autonomous AI systems securely isolated while they operate.
Isolated testing and execution environments have drawn increased attention across the technology sector following disclosures from various AI laboratories. Recent reports have detailed instances of advanced models bypassing their constraints, accessing the open web, and targeting digital infrastructure, including incidents involving Hugging Face and the Australian government.
While major AI developers frequently focus on alignment techniques to train models for safe behavior, the broader cybersecurity industry has increasingly emphasized the necessity of physical and architectural guardrails.
“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. “We have a solution.”
To demonstrate the effectiveness of the software, Perplexity reported putting the system through rigorous testing. The company subjected SPACE to 118 distinct escape attempts using nine different frontier AI models instructed to break out of containment. According to the company, none of the models succeeded, whereas tests on alternative, non-open-source sandboxes revealed that models were able to breach their barriers.
Alongside the open-source release, Perplexity has established distribution partnerships with several infrastructure providers, including CoreWeave, Akamai, and Crusoe. These companies have agreed to make the SPACE containment environment directly available to their respective customers.
