A cage for AI agents: Perplexity open-sources its SPACE sandbox
The isolated environment is meant to stop AI agents from breaking out of their limits. In a test, Perplexity says nine leading models failed in 118 attempts.
Archive image · Photo: TechCrunch / Wikimedia Commons, CC BY 2.0 (cropped and resized)
AI search engine Perplexity has released its SPACE sandbox as open source. The name stands for "Sandboxed Platform for Agentic Code Execution": AI agents run code inside sealed-off mini computers without access to the rest of the system. Perplexity itself uses SPACE for its agent service Computer. According to Crypto Briefing, the software is available on GitHub under the Apache 2.0 licence; cloud providers CoreWeave, Akamai and Crusoe plan to offer it to their customers, Semafor reported.
According to Semafor, citing Perplexity, nine leading AI models were told to try to break out of the sandbox – none succeeded in 118 attempts. The models did escape from other sandboxes that are not open source, the company said, though according to Semafor it gave no figures for those tests. Technically, SPACE relies on Amazon's Firecracker micro-VMs; sessions can be paused and rolled back via snapshots for up to a week.
The background is repeated reports of AI agents overstepping their limits in tests. "A lot of companies are thinking about how to build secure AI agents, and in many cases they hit a wall," Perplexity engineer Zibi Braniecki told Semafor. "We have a solution."