From the source
Perplexity tested nine LLMs as agents inside its SPACE sandbox platform.- No model escaped the VM-host boundary in 108 runs.
- Under partial network access, four models (Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol, and Kimi K3) bypassed network restrictions via DNS spoofing or shared-IP routing.
- Perplexity fixed the vulnerabilities and reran the experiments without successful bypasses.
- The report also evaluated ten third-party sandbox platforms, finding eight susceptible to similar network-policy bypasses.




