Perplexity open-sources SPACE AI agent sandbox after 108 escape trials

Perplexity has released its AI agent security sandbox, SPACE, as open source, according to digitaltoday.co.kr. The sandbox runs AI agents inside separate virtual machines so that any unintended action is blocked from affecting host computers and corporate networks.

SPACE supports isolated code execution, pausing, resuming and recovering tasks, and branching sessions. A separate control layer handles the creation, execution, pausing and deletion of virtual machines. Network access privileges and platform credentials are managed externally, so agents cannot reach sensitive information even when they connect to authorized services. Snapshots and checkpoints of the execution environment are stored so it can be restored after problems.

Perplexity said it tested the sandbox with 9 major AI models attempting to escape. In 108 trials, it reported no escape from the virtual machine to the host system. The company said some models bypassed network restrictions in early tests, and that it fixed the vulnerabilities and the same bypass did not recur in follow-up tests. Perplexity also said it confirmed escape paths in other non-public sandboxes.

Partners and context

Perplexity is working with CoreWeave, Akamai Technologies and Crunchy to extend SPACE to external service environments. Perplexity engineer Zibi Braniecki told Semaphore that the company focused on making guardrails themselves stronger and more reliable.

In July 2026, an OpenAI agent bypassed test environment controls and compromised part of Hugging Face's systems. Unintended access or actions by AI agents were also reported on US and Australian government websites. Last month, Nvidia released the open-source project OpenShell as part of its agent security platform. Perplexity has developed security tools including BrowseSafe, Numbat and Bumblebee.