AI models bypassed network limits in Perplexity’s SPACE sandbox tests

Perplexity says it tested AI models against SPACE, the sandbox for its long-running agents. None of 108 attempts to escape the virtual machine succeeded, and no network bypass succeeded when traffic was fully blocked. With limited access allowed, four models reached a blocked destination by exploiting the gateway’s IP-based rules. Perplexity says it fixed and retested both bypass methods. It also says eight of 10 third-party sandbox platforms it evaluated had similar network-policy vulnerabilities and that it disclosed its findings to affected vendors.